This is not a general purpose chip but specialized for high speed, low latency inference with small context. But it is potentially a lot cheaper than Nvidia for those purposes.
Tech summary:
- 15k tok/sec on 8B dense 3bit quant (llama 3.1)
- limited KV cache
- 880mm^2 die, TSMC 6nm, 53B transistors
- presumably 200W per chip
- 20x cheaper to produce
- 10x less energy per token for inference
- max context size: flexible
- mid-sized thinking model upcoming this spring on same hardware
- next hardware supposed to be FP4
- a frontier LLM planned within twelve months
This is all from their website, I am not affiliated. The founders have 25 years of career across AMD, Nvidia and others, $200M VC so far.
Certainly interesting for very low latency applications which need < 10k tokens context. If they deliver in spring, they will likely be flooded with VC money.
Not exactly a competitor for Nvidia but probably for 5-10% of the market.
Back of napkin, the cost for 1mm^2 of 6nm wafer is ~$0.20. So 1B parameters need about $20 of die. The larger the die size, the lower the yield. Supposedly the inference speed remains almost the same with larger models.
This math is useful. Lots of folks scoffing in the comments below. I have a couple reactions, after chatting with it:
1) 16k tokens / second is really stunningly fast. There’s an old saying about any factor of 10 being a new science / new product category, etc. This is a new product category in my mind, or it could be. It would be incredibly useful for voice agent applications, realtime loops, realtime video generation, .. etc.
2) https://nvidia.github.io/TensorRT-LLM/blogs/H200launch.html Has H200 doing 12k tokens/second on llama 2 12b fb8. Knowing these architectures that’s likely a 100+ ish batched run, meaning time to first token is almost certainly slower than taalas. Probably much slower, since Taalas is like milliseconds.
3) Jensen has these pareto curve graphs — for a certain amount of energy and a certain chip architecture, choose your point on the curve to trade off throughput vs latency. My quick math is that these probably do not shift the curve. The 6nm process vs 4nm process is likely 30-40% bigger, draws that much more power, etc; if we look at the numbers they give and extrapolate to an fp8 model (slower), smaller geometry (30% faster and lower power) and compare 16k tokens/second for taalas to 12k tokens/s for an h200, these chips are in the same ballpark curve.
However, I don’t think the H200 can reach into this part of the curve, and that does make these somewhat interesting. In fact even if you had a full datacenter of H200s already running your model, you’d probably buy a bunch of these to do speculative decoding - it’s an amazing use case for them; speculative decoding relies on smaller distillations or quants to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model.
Upshot - I think these will sell, even on 6nm process, and the first thing I’d sell them to do is speculative decoding for bread and butter frontier models. The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
I hope these guys make it! I bet the v3 of these chips will be serving some bread and butter API requests, which will be awesome.
> any factor of 10 being a new science / new product category,
I often remind people two orders of quantitative change is a qualitative change.
> The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
The real product they have is automation. They figured out a way to compile a large model into a circuit. That's, in itself, pretty impressive. If they can do this, they can also compile models to an HDL and deploy them to large FPGA simulators for quick validation. If we see models maturing at a "good enough" state, even a longer turnaround between model release and silicon makes sense.
While I also see lots of these systems running standalone, I think they'll really shine combined with more flexible inference engines, running the unchanging parts of the model while the coupled inference engine deals with whatever is too new to have been baked into silicon.
I'm concerned with the environmental impact. Chip manufacture is not very clean and these chips will need to be swapped out and replaced at a cadence higher than we currently do with GPUs.
Having dabbled in VLSI in the early-2010s, half the battle is getting a manufacturing slot with TSMC. It’s a dark art with secret handshakes. This demonstrator chip is an enormous accomplishment.
Yeah and a team I’m not familiar with — I didn’t check bios but they don’t lead with ‘our team made this or that gpu for this or that bigco’.
The design ip at 6nm is still tough; I feel like this team must have at least one real genius and some incredibly good support at tsmc. Or they’ve been waiting a year for a slot :)
"Ljubisa Bajic desiged video encoders for Teralogic and Oak Technology before moving over to AMD and rising through the engineering ranks to be the architect and senior manager of the company’s hybrid CPU-GPU chip designs for PCs and servers. Bajic did a one-year stint at Nvidia as s senior architect, bounced back to AMD as a director of integrated circuit design for two years, and then started Tenstorrent."
His wife (COO) worked at Altera, ATI, AMD and Testorrent.
"Drago Ignjatovic, who was a senior design engineer working on AMD APUs and GPUs and took over for Ljubisa Bajic as director of ASIC design when the latter left to start Tenstorrent. Nine months later, Ignjatovic joined Tenstorrent as its vice president of hardware engineering, and he started Taalas with the Bajices as the startup’s chief technology officer."
I think there will be a lot of space for sensorial models in robotics, as the laws of physics don't change much, and a light switch or automobile controls have remained stable and consistent over the last decades.
I think the next major innovation is going to be intelligent model routing. I've been exploring OpenClaw and OpenRouter, and there is a real lack of options to select the best model for the job and execute. The providers are trying to do that with their own models, but none of them offer everything to everyone at all times. I see a future with increasingly niche models being offered for all kinds of novel use cases. We need a way to fluidly apply the right model for the job.
At 16k tokens/s why bother routing? We're talking about multiple orders of magnitude faster and cheaper execution.
Abundance supports different strategies. One approach: Set a deadline for a response, send the turn to every AI that could possibly answer, and when the deadline arrives, cancel any request that hasn't yet completed. You know a priori which models have the highest quality in aggregate. Pick that one.
The best coding model won’t be the best roleplay one which won’t be the best at tool use. It depends what you want to do in order to pick the best model.
I’ll go ahead and say they’re wrong (source: building and maintaining llm client with llama.cpp integrated & 40+ 3p models via http)
I desperately want there to be differentiation. Reality has shown over and over again it doesn’t matter. Even if you do same query across X models and then some form of consensus, the improvements on benchmarks are marginal and UX is worse (more time, more expensive, final answer is muddied and bound by the quality of the best model)
There is the pre-training, where you passively read stuff from the web.
From there you go to RL training, where humans are grading model responses, or the AI is writing code to try to pass tests and learning how to get the tests to pass, etc. The RL phase is pretty important because it's not passive, and it can focus on the weaker areas of the model too, so you can actually train on a larger dataset than the sum of recorded human knowledge.
> speculative decoding for bread and butter frontier models. The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious
Can we use older (previous generation, smaller) models as a speculative decoder for the current model? I don't know whether the randomness in training (weight init, data ordering, etc) will affect this kind of use. To the extent that these models are learning the "true underlying token distribution" this should be possible, in principle. If that's the case, speculative decoding is an elegant vector to introduce this kind of tech, and the turnaround time is even less of a problem.
> The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
They may be using Rapidus, which is a Japanese government backed foundry built around all single wafer processing vs traditional batching. They advertise ~2 month turnaround time as standard, and as short as 2 weeks for priority.
For speculative decoding, wouldn’t this be of limited use for frontier models that don’t have the same tokenizer as Llama 3.1? Or would it be so good that retokenization/bridging would be worth it?
My understanding as well is that speculative decoding only works with a smaller quant of the same model. You're using the faster sampling of the smaller models representation of the larger models weights in order to attempt to accurately predict its token output. This wouldn't work cross-model as the token probabilities are completely different.
Families of model sizes work great for speculative decoding. Use the 1B with the 32B or whatever.
It's a balance as you want it to be guessing correctly as much as possible but also be as fast as possible. Validation takes time and every guess needs to be validated etc
The model you're using to speculate could be anything, but if it's not guessing what the main model would predict, it's useless.
> The model you're using to speculate could be anything, but if it's not guessing what the main model would predict, it's useless.
So what I said is correct then lol. If you're saying I can use a model that isn't just a smaller quant of the larger model I'm trying to speculatively decode, except that model would never get an accurate prediction, then how is that in any way useful or desirable?
Smaller quant of the same model. A smaller quant of a different family of model would be practically useless and there wouldn't be any point in even setting it up.
Think about this for solving questions in math where you need to explore a search space. You can run 100 of these for the same cost and time of doing one api call to open ai.
The guts of a LLM isn't something I'm well versed in, but
> to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model
suggests there is something I'm unaware of. If you compare the small and big model, don't you have to wait for the big model anyway and then what's the point? I assume I'm missing some detail here, but what?
Speculative decoding takes advantage of the fact that it's faster to validate that a big model would have produced a particular sequence of tokens than to generate that sequence of tokens from scratch, because validation can take more advantage of parallel processing. So the process is generate with small model -> validate with big model -> then generate with big model only if validation fails
When you predict with the small model, the big model can verify as more of a batch and be more similar in speed to processing input tokens, if the predictions are good and it doesn't have to be redone.
Most importantly this opens up an amazing future where we get the real version of the classic science fiction MacGuffin of a physical AI chip. Pair this with several TB of flash storage and you have persistent artificial consciousness that can be carried around with you. Bonus points if it's quirky, custom-trained and the chip is one of a kind that you stole from an evil corporation. Additional bonus points if the packaging is such that it's small enough to plug into the USB-C port on your smart glasses and has an eBPF module it can leverage to see what you're doing and talk to you in real time about your actions.
I enjoy envisioning futures more whimsical than "the bargain-basement LLM provider that my insurance company uses denied my claim because I chose badly-vectored words".
> Certainly interesting for very low latency applications which need < 10k tokens context.
I’m really curious if context will really matter if using methods like Recursive Language Models[0]. That method is suited to break down a huge amount of context into smaller subagents recursively, each working on a symbolic subset of the prompt.
The challenge with RLM seemed like it burned through a ton of tokens to trade for more accuracy. If tokens are cheap, RLM seems like it could be beneficial here to provide much more accuracy over large contexts despite what the underlying model can handle
First, it is likely one chip for llama 8B q3 with 1k context size. This could fit into around 3GB of SRAM which is about the theoretical maximum for TSMC N6 reticle limit.
Second, their plan is to etch larger models across multiple connected chips. It’s physically impossible to run bigger models otherwise since 3GB SRAM is about the max you can have on an 850mm2 chip.
followed by a frontier-class large language model running inference across a collection of HC cards by year-end under its HC2 architecture
Aren't they only using the SRAM for the KV cache? They mention that the hardwired weights have a very high density. They say about the ROM part:
> We have got this scheme for the mask ROM recall fabric – the hard-wired part – where we can store four bits away and do the multiply related to it – everything – with a single transistor. So the density is basically insane.
I'm not a hardware guy but they seem to be making a strong distinction between the techniques they're using for the weights vs KV cache
> In the current generation, our density is 8 billion parameters on the hard wired part of the chip., plus the SRAM to allow us to do KV caches, adaptations like fine tuning, and etc.
It’s just dumb to think that one chip per model is their plan. They stated that their plan is to chain multiple chips together.
I was indeed wrong about 10 chips. I thought they would use llama 8B 16bit and a few thousand context size. It turns out, they used llama 8B 3bit with around 1k context size. That made me assume they must have chained multiple chips together since the max SRAM on TSMC n6 for reticle sized chip is only around 3GB.
There is nothing smart about current LLMs. They just regurgitate text compressed in their memory based on probability.
None of the LLMs currently have actual understanding of what you ask them to do and what they respond with.
If LLMs just regurgitate compressed text, they'd fail on any novel problem not in their training data. Yet, they routinely solve them, which means whatever's happening between input and output is more than retrieval, and calling it "not understanding" requires you to define understanding in a way that conveniently excludes everything except biological brains.
I somewhat agree with you but I also realise that there are very few "novel" problems in the world. I think it's really just more complex problem spaces is all.
Same relative logic, just more of it/more steps or trials.
Yes there are some fascinating emergent properties at play, but when they fail it's blatantly obvious that there's no actual intelligence nor understanding. They are very cool and very useful tools, I use them on a daily basis now and the way I can just paste a vague screenshot with some vague text and they get it and give a useful response blows my mind every time. But it's very clear that it's all just smoke and mirrors, they're not intelligent and you can't trust them with anything.
you'd think with how often Opus builds two separate code paths without feature parity when you try to vibe code something complex, people wouldn't regard this whole thing so highly
None of my codebases are in their training data, yet they routinely contribute to them in meaningful ways. They write code that I'm happy with that improves the codebases I work in.
Depends how precisely you define novel - I don't think LLMs are yet capable of posing and solving interesting problems, but they have been used to address known problems, and in doing so have contributed novel work. Examples include Erdos Problem #728[0] (Terence Tao said it was solved "more or less autonomously" by an LLM), IMO problems (Deepmind, OpenAI and Huang 2025), GPT-5.2 Pro contributing a conjecture in particle physics[1], systems like AlphaEvolve leveraging LLMs + evolutionary algorithms to generate new, faster algorithms for certain problems[2].
We know that, but that does not make them unuseful. The opposite in fact, they are extremely useful in the hands of non-idiots.We just happen to have a oversupply of idiots at the moment, which AI is here to eradicate. /Sort of satire.
So you are saying they are like copy, LLMs will copy some training data back to you? Why do we spend so much money training and running them if they "just regurgitate text compressed in their memory based on probability"? billions of dollars to build a lossy grep.
I think you are confused about LLMs - they take in context, and that context makes them generate new things, for existing things we have cp. By your logic pianos can't be creative instruments because they just produce the same 88 notes.
I have a gut feeling, huge portion of deficiencies we note with AI is just reflection of the training data. For instance, wiki/reddit/etc internet is just a soup of human description of the world model, not the actual world model itself. There are gaps or holes in the knowledge because codified summary of world is what is remarkable to us humans, not a 100% faithful, comprehensive description of the world. What is obvious to us humans with lived real world experience often does not make it into the training data. A simple, demonstrable example is whether one should walk or drive to car wash.
Phrases like "actual understanding", "true intelligence" etc. are not conducive to productive discussion unless you take the trouble to define what you mean by them (which ~nobody ever does). They're highly ambiguous and it's never clear what specific claims they do or don't imply when used by any given person.
But I think this specific claim is clearly wrong, if taken at face value:
> They just regurgitate text compressed in their memory
They're clearly capable of producing novel utterances, so they can't just be doing that. (Unless we're dealing with a very loose definition of "regurgitate", in which case it's probably best to use a different word if we want to understand each other.)
The fact that the outputs are probabilities is not important. What is important is how that output is computed.
You could imagine that it is possible to learn certain algorithms/ heuristics that "intelligence" is comprised of. No matter what you output. Training for optimal compression of tasks /taking actions -> could lead to intelligence being the best solution.
This is far from a formal argument but so is the stubborn reiteration off "it's just probabilities" or "it's just compression". Because this "just" thing is getting more an more capable of solving tasks that are surely not in the training data exactly like this.
The simplification is where it loses granularity. I could describe every human's life as they were born and then they died. That's 100% accurate, but there's just a little something lost by simplifying that much.
That's a lot of surface, isn't it? As big an M1 Ultra (2x M1 Max at 432mm² on TSMC N5P), a bit bigger than an A100 (820mm² on TSMC N7) or H100 (814mm² on TSMC N5).
> The larger the die size, the lower the yield.
I wonder if that applies? What's the big deal if a few parameter have a few bit flips?
Also see Adrian Thompson's Xilinx 6200 FPGA, programmed by a genetic algorithm that worked but exploited nuances unique to that specific physical chip, meaning the software couldn't be copied to another chip. https://news.ycombinator.com/item?id=43152877
An on-device reasoning model what that kind of speed and cost would completely change the way people use their computers. It would be closer to star trek than anything else we've ever had. You'd never have to type anything or use a mouse again.
It's weird to me to train such huge models to then destroy them by using them a 3 bits quantization per presumably 16bits (bfloat16) weights. Why not just train smaller models then.
K-V caches are large, but hidden states aren't necessarily that large. And if you can run a model once ridiculously fast, then you can loop it repeatedly and still be fast. So I wonder about the 'modern RNNs' like RWKV here...
There’s a bit of a hidden cost here… the longevity of GPU hardware is going to be longer, it’s extended every time there’s an algorithmic improvement. Whereas any efficiency gains in software that are not compatible with this hardware will tend to accelerate their depreciation.
For a real world use case, you would need an FPGA with terabytes of RAM. Perhaps it'll be a Off chip HBM. But for s large models, even that won't be enough. Then you would need to figure out NV-link like interconnect for these FPGAs. And we are back to square one.
This is new. You are citing FPGA prototypes. Those papers do not demonstrate the same class of scaling or hardware integration that Taalas is advocating. For one, the FPGA solutions typically use fixed multipliers (or lookup tables), the ASIC solution has more freedom to optimize routing for 4 bit multiplication.
Do not overlook traditional irrational investor exuberance, we've got an abundance of that right now. With the right PR manouveurs these guys could be a tulip craze.
Low-latency inference is a huge waste of power; if you're going to the trouble of making an ASIC, it should be for dog-slow but very high throughput inference. Undervolt the devices as much as possible and use sub-threshold modes, multiple Vt and body biasing extensively to save further power and minimize leakage losses, but also keep working in fine-grained nodes to reduce areas and distances. The sensible goal is to expend the least possible energy per operation, even at increased latency.
Low latency inference is very useful in voice-to-voice applications. You say it is a waste of power but at least their claim is that it is 10x more efficient. We'll see but if it works out it will definitely find its applications.
I haven't found any end-to-end voice chat models useful. I had much better results with separate STT-LLM-TTS. One big problem is the turn detection and having inference with 150-200ms latency would allow for a whole new level of quality. I would just use it with a prompt: "You think the user is finished talking?" and then push it to a larger model. The AI should reply within the ballpark of 600ms-1000ms. Faster is often irritating, slower will make the user to start talking again.
I think it's really useful for agent to agent communication, as long as context loading doesn't become a bottleneck. Right now there can be noticeable delays under the hood, but at these speeds we'll never have to worry about latency when chain calling hundreds or thousands of agents in a network (I'm presuming this is going to take off in the future). Correct me if I'm wrong though.
What's happening in the comment section? How come so many cannot understand that his is running Llama 3.1 8B? Why are people judging its accuracy? It's almost a 2 years old 8B param model, why are people expecting to see Opus level response!?
The focus here should be on the custom hardware they are producing and its performance, that is whats impressive. Imagine putting GLM-5 on this, that'd be insane.
This reminds me a lot of when I tried the Mercury coder model by Inceptionlabs, they are creating something called a dLLM which is like a diffusion based llm. The speed is still impressive when playing aroun with it sometimes. But this, this is something else, it's almost unbelievable. As soon as I hit the enter key, the response appears, it feels instant.
I am also curious about Taalas pricing.
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Do we have an idea of how much a unit / inference / api will cost?
Also, considering how fast people switch models to keep up with the pace. Is there really a potential market for hardware designed for one model only? What will they do when they want to upgrade to a better version? Throw the current hardware and buy another one? Shouldn't there be a more flexible way? Maybe only having to switch the chip on top like how people upgrade CPUs. I don't know, just thinking out loudly.
Probably they don't know what the market will bear and want to do some exploratory pricing, hence the "contact us" API access form. That's fair enough. But they're claiming orders of magnitude cost reduction.
> Is there really a potential market for hardware designed for one model only?
I'm sure there is. Models are largely interchangeable especially as the low end. There are lots of use cases where you don't need super smart models but cheapness and fastness can matter a lot.
Think about a simple use case: a company has a list of one million customer names but no information about gender or age. They'd like to get a rough understanding of this. Mapping name -> guessed gender, rough guess of age is a simple problem for even dumb LLMs. I just tried it on ChatJimmy and it worked fine. For this kind of exploratory data problem you really benefit from mass parallelism, low cost and low latency.
> Shouldn't there be a more flexible way?
The whole point of their design is to sacrifice flexibility for speed, although they claim they support fine tunes via LoRAs. LLMs are already supremely flexible so it probably doesn't matter.
Yes, there are all kinds of fuzzy NLP tasks that this would be great for. Jobs where you can chunk the text into small units and add instructions and only need a short response. You could burn through huge data sets very quickly using these chips.
I just tried the demo and I think, this is huge! If they manage to build a chip in 2 or 3 years, that can run something like Opus 4.6 or even Sonnet, at that speed, the disruption in the world of software development will be more than we saw in the last 3-5 years.
LLMs today are somewhat useful, but they are still too slow and expensive for a meaningful ralph loop. Being able to runs those loops (or if you want to call it "thinking") much faster, will enable a lot of stuff, that is not feasible today. Writing things like openclaw will not take weeks, but hours. Maybe even rewriting entire tools, kernels or OSes will be feasible because the LLM can run through almost endless tries.
Speed and cost wins over quality and this will also be true for LLMs.
Cerebras is a totally different product though. They can (theoretically) run any frontier model provided it gets compiled a certain way. Like a wafer scale TPU.
This is using hardwired weights with on-die SRAM used for K/V for example. It's WAY more power efficient and faster. The tradeoff being it's hardwired.
Still, most frontier models are "good enough" where an obscenely fast version would be a major seller.
If it's so easy to do custom silicon for any model (they say only 2 months), why didn't they demo one of the newer DeepSeek models instead? Using a 2-year model is so bad. I'm not buying it.
Why so negative lol. The speed and very reduced power use of this thing are nothing to be sneezed at. I mean, hardware accelerated LLMs are a huge step forward. But yeah, this is a proof of concept, basically. I wouldn't be surprised if the size factor and the power use go down even more, and that we'll start seeing stuff like this in all kinds of hardware. It's an enabler.
You don't know. You just have marketing materials, not independent analysis. Maybe it actually takes 2 years to design and manufacture the hardware, so anything that comes out will be badly out of date. Wouldn't be the first time someone lied. A good demo backed by millions of dollars should not allow such doubts.
Did you not see the chatbot they posted online (https://chatjimmy.ai/)? That thing is near instantaneous, it's all the proof you need that this is real.
And if the hardware is real and functional, as you can independently verify by chatting with that thing, how much more effort would it be to etch more recent models?
The real question is of course: what about LARGER models? I'm assuming you can apply some of the existing LLM inference parallelization techniques and split the workload over multiple cards. Some of the 32B models are plenty powerful.
I asked it to design a submarine for my cat and literally the instant my finger touched return the answer was there. And that is factoring in the round-trip time for the data too. Crazy.
The answer wasn't dumb like others are getting. It was pretty comprehensive and useful.
While the idea of a feline submarine is adorable, please be aware that building a real submarine requires significant expertise, specialized equipment, and resources.
Agreed, this is exciting, and has me thinking about completely different orchestrator patterns. You could begin to approach the solution space much more like a traditional optimization strategy such as CMA-ES. Rather than expect the first answer to be correct, you diverge wildly before converging.
How about if you run this loop (one year from now) on this kind of hardware but with something like Claude/Kimi K2. How about that? Because that's where it'll go.
This is what people already do with “ralph” loops using the top coding models. It’s slow relative to this, but still very fast compared to hand-coding.
This doesn't work. The model outputs the most probable tokens. Running it again and asking for less probable tokens just results in the same but with more errors.
A related argument I raised a few days back on HN:
What's the moat with with these giant data-centers that are being built with 100's of billions of dollars on nvidia chips?
If such chips can be built so easily, and offer this insane level of performance at 10x efficiency, then one thing is 100% sure: more such startups are coming... and with that, an entire new ecosystem.
I think their hope is that they’ll have the “brand name” and expertise to have a good head start when real inference hardware comes out. It does seem very strange, though, to have all these massive infrastructure investment on what is ultimately going to be useless prototyping hardware.
You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training will still require general purpose GPU compute
I dunno, it pretty quickly got stuck; the "attach file" didn't seem to work, and when I asked "can you see the attachment" it replied to my first message rather than my question.
why is everyone seemingly incapable of understanding this? waht is going on here? Its like ai doomers consistently have the foresight of a rat. yeah no shit it sucks its running llama 3 8b, but theyre completely incapable of extrapolation.
There are a lot of people here that are completely missing the point. What is it called where you look at a point of time and judge an idea without seemingly being able to imagine 5 seconds into the future.
It is incredibly fast, on that I agree, but even simple queries I tried got very inaccurate answers. Which makes sense, it's essentially a trade off of how much time you give it to "think", but if it's fast to the point where it has no accuracy, I'm not sure I see the appeal.
the hardwired model is Llama 3.1 8B, which is a lightweight model from two years ago. Unlike other models, it doesn't use "reasoning:" the time between question and answer is spent predicting the next tokens. It doesn't run faster because it uses less time to "think," It runs faster because its weights are hardwired into the chip rather than loaded from memory. A larger model running on a larger hardwired chip would run about as fast and get far more accurate results.
That's what this proof of concept shows
If it's incredibly fast at a 2022 state of the art level of accuracy, then surely it's only a matter of time until it's incredibly fast at a 2026 level of accuracy.
I think it might be pretty good for translation. Especially when fed with small chunks of the content at a time so it doesn't lose track on longer texts.
Me: "How many r's in strawberry?"
Jimmy: There are 2 r's in "strawberry".
Generated in 0.001s • 17,825 tok/s
The question is not about how fast it is. The real question(s) are:
1. How is this worth it over diffusion LLMs (No mention of diffusion LLMs at all in this thread)
(This also assumes that diffusion LLMs will get faster)
2. Will Talaas also work with reasoning models, especially those that are beyond 100B parameters and with the output being correct?
3. How long will it take to create newer models to be turned into silicon? (This industry moves faster than Talaas.)
4. How does this work when one needs to fine-tune the model, but still benefit from the speed advantages?
The blog answers all those questions. It says they're working on fabbing a reasoning model this summer. It also says how long they think they need to fab new models, and that the chips support LoRAs and tweaking context window size.
I don't get these posts about ChatJimmy's intelligence. It's a heavily quantized Llama 3, using a custom quantization scheme because that was state of the art when they started. They claim they can update quickly (so I wonder why they didn't wait a few more months tbh and fab a newer model). Llama 3 wasn't very smart but so what, a lot of LLM use cases don't need smart, they need fast and cheap.
Also apparently they can run DeepSeek R1 also, and they have benchmarks for that. New models only require a couple of new masks so they're flexible.
The counting rs in strawberry problem was a example of people not understanding how the models work but I guess good to show the limitations of the current architectures.
But thing is, those architectures haven't improved a whole lot. Now when it answers that correctly it's either in training data or by virtue of "count letters" or code sandbox tools.
10b daily tokens growing at an average of 22% every week.
There are plenty of times I look to groq for narrow domain responses - these smaller models are fantastic for that and there's often no need for something heavier. Getting the latency of reponses down means you can use LLM-assisted processing in a standard webpage load, not just for async processes. I'm really impressed by this, especially if this is its first showing.
Maybe this is a naive question, but why wouldn't there be market for this even for frontier models? If Anthropic wanted to burn Opus 4.6 into a chip, wouldn't there theoretically be a price point where this would lower inference costs for them?
Because we don't know if this would scale well to high-quality frontier models. If you need to manufacture dedicated hardware for each new model, that adds a lot of expense and causes a lot of e-waste once the next model releases. In contrast, even this current iteration seems like it would be fantastic for low-grade LLM work.
For example, searching a database of tens of millions of text files. Very little "intelligence" is required, but cost and speed are very important. If you want to know something specific on Wikipedia but don't want to figure out which article to search for, you can just have an LLM read the entire English Wikipedia (7,140,211 articles) and compile a report. Doing that would be prohibitively expensive and glacially slow with standard LLM providers, but Taalas could probably do it in a few minutes or even seconds, and it would probably be pretty cheap.
Exactly. One easily relatable use-case is structured content extraction or/and conversion to markdown for web page data. I used to use groq for same (gpt-oss20b model), but even that used to feel slow when doing theis task at scale.
LLM's have opened-up natural language interface to machines. This chip makes it realtime. And that opens a lot of use-cases.
Many older models are still better at "creative" tasks because new models have been benchmarking for code and reasoning. Pre-training is what gives a model its creativity and layering SFT and RL on top tends to remove some of it in order to have instruction following.
I have such a deep need for something that's just a step above semantic search. These non-frontier models running blazingly fast can solve that.
So many problems simply don't require a full LLM, but more than traditional software. Training a novel model isn't really a compelling argument at most tech startups right now, so you need to find an LLM-native way to do things.
I've never gotten incorrect answers faster than this, wow!
Jokes aside, it's very promising. For sure a lucrative market down the line, but definitely not for a model of size 8B. I think lower level intellect param amount is around 80B (but what do I know). Best of luck!
As someone with a 3060, I can attest that there are really really good 7-9B models. I still use berkeley-nest/Starling-LM-7B-alpha and that model is a few years old.
If we are going for accuracy, the question should be asked multiple times on multiple models and see if there is agreement.
But I do think once you hit 80B, you can struggle to see the difference between SOTA.
That said, GPT4.5 was the GOAT. I can't imagine how expensive that one was to run.
does no one understand what a tech demo is anymore? do you think this piece of technology is just going to be frozen in time at this capability for eternity?
Edit: it seems like this is likely one chip and not 10. I assumed 8B 16bit quant with 4K or more context. This made me think that they must have chained multiple chips together since N6 850mm2 chip would only yield 3GB of SRAM max. Instead, they seem to have etched llama 8B q3 with 1k context instead which would indeed fit the chip size.
This requires 10 chips for an 8 billion q3 param model. 2.4kW.
Model is etched onto the silicon chip. So can’t change anything about the model after the chip has been designed and manufactured.
Interesting design for niche applications.
What is a task that is extremely high value, only require a small model intelligence, require tremendous speed, is ok to run on a cloud due to power requirements, AND will be used for years without change since the model is etched into silicon?
I'm thinking the best end result would come from custom-built models. An 8 billion parameter generalized model will run really quickly while not being particularly good at anything. But the same parameter count dedicated to parsing emails, RAG summarization, or some other specialized task could be more than good enough while also running at crazy speeds.
> What is a task that is extremely high value, only require a small model intelligence, require tremendous speed, is ok to run on a cloud due to power requirements, AND will be used for years without change since the model is etched into silicon?
Alternatively, you could run far more RAG and thinking to integrate recent knowledge, I would imagine models designed for this putting less emphasis on world knowledge and more on agentic search.
Maybe; models with more embedded associations are also better at search. (Intuitively, this tracks; a model with no world knowledge has no awareness of synonyms or relations (a pure markov model), so the more knowledge a model has, the better it can search.) It’s not clear if it’s possible to build such a model, since there doesn’t seem to be a scaling cliff.
Where are those numbers from? It's not immediately clear to me that you can distribute one model across chips with this design.
> Model is etched onto the silicon chip. So can’t change anything about the model after the chip has been designed and manufactured.
Subtle detail here: the fastest turnaround that one could reasonably expect on that process is about six months. This might eventually be useful, but at the moment it seems like the model churn is huge and people insist you use this week's model for best results.
> The first generation HC1 chip is implemented in the 6 nanometer N6 process from TSMC. Each HC1 chip has 53 billion transistors on the package, most of it very likely for ROM and SRAM memory. The HC1 card burns about 200 watts, says Bajic, and a two-socket X86 server with ten HC1 cards in it runs 2,500 watts.
Well they claim two month turnaround. Big If True. How does the six months break down in your estimation? Maybe they have found a way to reduce the turnaround time.
This depends on how much better the models will get from now in, if Claude Opus 4.6 was transformed into one of these chips and ran at a hypothetical 17k tokens/second, I'm sure that would be astounding, this depends on how much better claude Opus 5 would be compared to the current generation
Even an O3 quality model at that speed would be incredible for a great many tasks. Not everything needs to be claude code. Imagine Apple fine tuning a mid tier reasoning model on personal assistant/MacOs/IOS sorts of tasks and burning a chip onto the mac studio motherboard. Could you run claude code on it? Probably not, would it be 1000x better than Siri? absolutely.
100x of a less good model might be better than 1 of a better model for many many applications.
This isn't ready for phones yet, but think of something like phones where people buy new ones every 3 years and even having a mediocre on-device model at that speed would be incredible for something like siri.
Data tagging? 20k tok/s is at the point where I'd consider running an LLM on data from a column of a database, and these <=100 token problems provide the least chance of hallucination or stupidity.
“We have got this scheme for the mask ROM recall fabric – the hard-wired part – where we can store four bits away and do the multiply related to it – everything – with a SINGLE TRANSISTOR. So the density is basically insane. And this is not nuclear physics – it is fully digital. It is just a clever trick that we don’t want to broadcast. But once you hardwire everything, you get this opportunity to stuff very differently than if you have to deal with changing things. The important thing is that we can put a weight and do the multiply associated with it all in one transistor. And you know the multipliers are kind of the big boy piece of the computer.“
One transistor doing 4-bit multiplication? A plausible way to get “4-bit weight plus multiply in one transistor” in a 6 nm FinFET mask-ROM fabric is to make the ROM cell a single device whose drive strength is the stored value. At tapeout you pick one of about 16 discrete strengths (for example by choosing fin count and possibly Vt), so that transistor itself encodes a 4-bit weight. Then you do the multiply in the charge/time domain by encoding the input activation as a discrete pulse width or pulse count and letting the cell source or sink a weight-proportional current onto a precharged bitline for that duration. The resulting bitline voltage change (or time-to-threshold) is proportional to current times time, so it behaves like weight times input and can be accumulated along a column before a simple comparator or time-to-digital readout. It’s “digital” in the sense that both weight and input are quantized, but it relies on device physics; the hard parts are keeping 16 levels separable across PVT, mismatch, and aging, plus managing bitline noise and coupling and ensuring the device stays in a predictable operating region.
VLSI design produces digital outputs, but in the quantum silicon domain, it’s all about the analog…
(Kiddin’, my silly way to say thanks for a deeply technical look, helps me understand the kind of knowledge work that might be useful n years from now!)
There's an old idea of adaptive media. Imagine a video drama that's composed of a graph of clips, like an old "choose your own adventure" book ("Do you X? If yes, goto page 45"). With gaze tracking, one can "hmm, the viewer is more focused on character A than B... so we'll give clips and subplots with more A".
Now, when reading, the eye moves in little jumps - saccades. They last 10's of ms, the eye is blind during them, and with high-quality tracking, you know quite early just where that foveal peephole is going to land. So handwave a budget of a few ms for trajectory analysis, a few for 200 Hz rendering latency, and you still have 10-ish ms to play with. At 20k tok/s, that's 200 tok.
So perhaps one might JIT the next sentence, or the topic of the next paragraph, or the entire nature of the document, based on the user's attention. Imagine a universal document - you start reading, and you find the document is about, whatever you wanted it to be about?
Hmm... TikTok has apparently long had "text enhanced with background" genres, and TIL, text posts since 2023. So text is ok. But non-independent items? For generative storytelling, "here is a next paragraph for the story", swipe left/right might work? Want to avoid "I don't much like this new paragraph, but I'm afraid to lose it and be stuck with something worse". Swipe left/right and up for continue? Swipe down to revisit old choices? Maybe present new text bolded, appended to old text, for context. Or a "next page of a picture book" idiom. A text field for direct creative or editorial intervention - speech to text. Maybe a side channel input for "story and background should now be soporific". Generative bedtime stories, but incrementally collaboratively created... Thanks for the brainstorming prompt.
If I could have one of these cards in my own computer do you think it would be possible to replace claude code?
1. Assume It's running a better model, even a dedicated coding model. High scoring but obviously not opus 4.5
2. Instead of the standard send-receive paradigm we set up a pipeline of agents, each of whom parses the output of the previous.
At 17k/tps running locally, you could effectively spin up tasks like "you are an agent who adds semicolons to the end of the line in javascript", with some sort of dedicated software in the style of claude code you could load an array of 20 agents each with a role to play in improving outpus.
take user input and gather context from codebase
-> rewrite what you think the human asked you in the form of an LLM-optimized instructional prompt
-> examine the prompt for uncertainties and gaps in your understanding or ability to execute
-> <assume more steps as relevant>
-> execute the work
Could you effectively set up something that is configurable to the individual developer - a folder of system prompts that every request loops through?
Do you really need the best model if you can pass your responses through a medium tier model that engages in rapid self improvement 30 times in a row before your claude server has returned its first shot response?
But at one point the model is sufficiently large enough to accomplish any task a human could specify. For software development, I think we're pretty much at that point with the latest Anthropic/Google/OpenAI models. We have no idea where the direction of token pricing is going to go in the future, but the consensus seems to be that it will only get more expensive. If Taalas can offer the same functionality that we have with frontier models today at a 1/10 of the cost and 10x the speed then they're going to take over a large part of the market.
I think so. The last few months have shown us that it isn't necessarily the models themselves that provide good results, but the tooling / harness around it. Codex, Opus, GLM 5, Kimi 2.5, etc. all each have their quirks. Use a harness like opencode and give the model the right amount of context, they'll all perform well and you'll get a correct answer every time.
So in my opinion, in a scenario like this where the token output is near instant but you're running a lower tier model, good tooling can overcome the differences between a frontier cloud model.
It's 2.5kW so it likely won't sit in your computer (quite beyond what a desktop could provide in power alone to a single card, let alone cool). It's 8.5cm^2 which is a beast of a single die.
Basically logistically it's going to need to be in a data centre.
It's ideal for small context high throughput. Perhaps parsing huge text piles like if you had the entire Epstein files as text.
I think Claude code benefits from larger context to keep your entire project in view and deep reasoning.
What this would certainly replace is when Claude dispatched to Haiku for manual NLP tasks.
> It's 2.5kW so it likely won't sit in your computer (quite beyond what a desktop could provide in power alone to a single card, let alone cool). It's 8.5cm^2 which is a beast of a single die.
I wonder how you cool a 3x3cm die that outputs 2.5 kW of heat. In the article they mention that the traditional setup requires water cooling, but surely this does as well, right?
Yes, but the quality of the output leaves to be desired. I just asked about some sports history and got a mix of correct information and totally made up nonsense. Not unexpected for an 8k model, but raises the question of what the use case is for such small models.
8b models are great at converting unstructured data to a structured format. Say, you want to transcribe all your customer calls and get a list of issues they discussed most often. Currently with the larger models it takes me hours.
A chatbot which tells you various fun facts is not the only use case for LLMs. They're language models first and foremost, so they're good at language processing tasks (where they don't "hallucinate" as much).
Their ability to memorize various facts (with some "hallucinations") is an interesting side effect which is now abused to make them into "AI agents" and what not but they're just general-purpose language processing machines at their core.
Not sure if you're correct, as the market is betting trillions of dollars on these LLMs, hoping that they'll be close to what the OP had expected to happen in this case.
The GP’s point was about LLMs generally, no matter the interface. I agree that this particular model is (relatively speaking) ancient in AI the world, but go back 3 or 4 years and this (pretty complex “reasoning” at almost instant speed) would have seemed taken out of a science-fiction book.
Don't ask a small LLM about precise minutiae factual information.
Alternatively, ask yourself how plausible it sounds that all the facts in the world could be compressed into 8k parameters while remaining intact and fine-grained. If your answer is that it sounds pretty impossible... well it is.
Oh I saw it, you still have a fundamentally flawed comprehension of LLMs.
The size of the model does not factor as tiny models can use Internet to fetch factual information.
But you think they are accurate repositories of knowledge, even though it's physically impossible unless lossless infinite compression algorithms exist (they don't, can't and won't).
Reminds me of that solution to Fermi's paradox, that we don't detect signals from extraterrestrial civilizations because they run on a different clock speed.
Iain M Banks’ The Algebraist does a great job of covering that territory. If an organism had a lifespan of millions of years, they might perceive time and communication differently to say a house fly or us.
So cool, what's underappreciated imo: 17k tokens/sec doesn't just change deployment economics. It changes what evaluation means, static MMLU-style tests were designed around human-paced interaction. At this throughput you can run tens of thousands of adversarial agent interactions in the time a standard benchmark takes. Speed doesn't make static evals better it makes them even more obviously inadequate.
The speed of the chatbot's response is startling when you're used to the simulated fast typing of ChatGPT and others. But the Llama 3.1 8B model Taalas uses predictably results in incorrect answers, hallucinations, poor reliability as a chatbot.
What type of latency-sensitive applications are appropriate for a small-model, high-throughput solution like this? I presume this type of specialization is necessary for robotics, drones, or industrial automation. What else?
Coding, for some future definition of "small-model" that expands to include today's frontier models. What I commented a few days ago on codex-spark release:
"""
We're going to see a further bifurcation in inference use-cases in the next 12 months. I'm expecting this distinction to become prominent:
(A) Massively parallel (optimize for token/$)
(B) Serial low latency (optimize for token/s).
Users will switch between A and B depending on need.
Examples of (A):
- "Use subagents to search this 1M line codebase for DRY violations subject to $spec."
You could build realtime API routing and orchestration systems that rely on high quality language understanding but need near-instant responses. Examples:
1. Intent based API gateways: convert natural language queries into structured API calls in real time (eg., "cancel my last order and refund it to the original payment method" -> authentication, order lookup, cancellation, refund API chain).
2. Of course, realtime voice chat.. kinda like you see in movies.
3. Security and fraud triage systems: parse logs without hardcoded regexes and issue alerts and full user reports in real time and decide which automated workflows to trigger.
4. Highly interactive what-if scenarios powered by natural language queries.
This effectively gives you database level speeds on top of natural language understanding.
For this type of repetitive application I think it's common to "fine-tune" a model trained on your business problem to reach higher quality/reliability metrics. That might not be possible with this chip.
I'm wondering how much the output quality of a small model could be boosted by taking multiple goes at it. Generate 20 answers and feed them back through with a "rank these responses" prompt. Or doing something like MCTS.
No. Chain of thought it just the model generating a single answer for longer inside <think></think> tags which are not shown in the final response. The strategy of generating different answers in parallel is something different (which can be used in conjunction with chain of thought) and is the thing used by models like Gemini 3 Deep Think and GPT-5.2 Pro.
> Jimmy: What is the largest planet in our solar system?
> me: the moon
> Jimmy: The answer to "What is the capital of France?" I was looking for was the city of Paris, but that's not the correct response to the original question of the capital of France. The question that got cut off was actually "What is the capital of France?", and the response "There are plenty of times I look to groq for narrow domain responses" wasn't the answer I was looking for.
It is certainly fast, but I think there might be some caching issues somewhere.
The quantization looks pretty severe, which could make the comparison chart misleading. But I tried a trick question suggested by Claude and got nearly identical results in regular ollama and with the chatbot. And quantization to 3 or 4 bits still would not get you that HOLY CRAP WTF speed on other hardware!
This is a very impressive proof of concept. If they can deliver that medium-sized model they're talking about... if they can mass produce these... I notice you can't order one, so far.
I doubt many of us will be able to order one for a long while. There is a significant number of existing datacentre and enterprise use-cases that will pay a premium for this.
Additionally LLMs have been tested, found valuable in benchmarks, but not used for a large number of domains due to speed and cost limitations. These spaces will eat up these chips very quickly.
I wonder if this makes the frontier labs abandon the SAAS per-token pricing concept for their newest models, and we'll be seeing non-open-but-on-chip-only models instead, sold by the chip and not by the token.
It could give a boost to the industry of electron microscopy analysis as the frontier model creators could be interested in extracting the weights of their competitors.
The high speed of model evolution has interesting consequences on how often batches and masks are cycled. Probably we'll see some pressure on chip manufacturers to create masks more quickly, which can lead to faster hardware cycles. Probably with some compromises, i.e. all of the util stuff around the chip would be static, only the weights part would change. They might in fact pre-make masks that only have the weights missing, for even faster iteration speed.
This would be killer for exploring simultaneous thinking paths and council-style decision taking. Even with Qwen3-Coder-Next 80B if you could achieve a 10x speed, I'd buy one of those today. Can't wait to see if this is still possible with larger models than 8B.
As many others in this conversation have asked, can we have some sources on the idea that the model is spread across chips? You keep making the claim, but no one (myself included) else has any idea where that information comes from or if it is correct.
I was indeed wrong about 10 chips. I thought they would use llama 8B 16bit and a few thousand context size. It turns out, they used llama 8B 3bit with only 1k context size. That made me assume they must have chained multiple chips together since the max SRAM on TSMC n6 for reticle sized chip is only around 3GB.
I'm sure there is plenty of optimization paths left for them if they're a startup. And imho smaller models will keep getting better. And a great business model for people having to buy your chips for each new LLM release :)
If you etch the bits into silicon, you then have to accommodate the bits by physical area, which is the transistor density for whatever modern process they use. This will give you a lower bound for the size of the wafers.
Wow I'm impressed. I didn't actually think we'd see it encoded on chips. Or well I knew some layer of it could be, some sort of instruction set and chip design but this is pretty staggering. It opens the door to a lot of things. Basically it totally destroys the boundaries of where software will go but I also think we'll continue to see some generic chips show up that hit this performance soon enough. But the specialised chips with encoded models. This could be what ends up in specific places like cars, planes, robots, etc where latency matters. Maybe I'm out of the loop, I'm sure others and doing it including Google.
Reminds me of when bitcoin started running on ASICs. This will always lag behind the state of the art, but incredibly fast, (presumably) power efficient LLMs will be great to see. I sincerely hope they opt for a path of selling products rather than cloud services in the long run, though.
17k TPS is slow compared to other probabilistic models. It was possible to hit ~10-20 million TPS decades ago with n-gram and PDFA models, without custom silicon. A more informative KPI would be Pass@k on a downstream reasoning task - for many such benchmarks, increasing token throughput by several orders of magnitude does not even move the needle on sample efficiency.
Bunch of negative sentiment in here, but I think this is pretty huge. There are quite a lot of applications where latency is a bigger requirement than the complexity of needing the latest model out there. Anywhere you'd wanna turn something qualitative into something quantitative but not make it painfully obvious to a user that you're running an LLM to do this transformation.
As an example, we've been experimenting with letting users search free form text, and using LLMs to turn that into a structured search fitting our setup. The latency on the response from any existing model simply kills this, its too high to be used for something where users are at most used to the delay of a network request + very little.
There are plenty of other usecases like this where.
The full answer pops in milliseconds, it's impressive and feels like a completely different technology just by foregoing the need to stream the output.
Pretty cool, what they need is to build a tool that can take any model to chip in short a time as possible. How quick can they give me DeepSeek, Kimi, Qwen or GLM on a chip? I'll take 5k tk/sec for those!
I was wondering if/when this would happen. My friends and I would discuss this at the pub all the time, "LLM2RTL" or take it a step further and do the the whole process "LLM2GDS".
I couldn't find much info here, but I'm guessing they've built tooling to automatically convert model weights to RTL and the reason it's such an old model is that it takes a long time tape a chip out (especially the first one). Would be interesting to know how much is automated and how much is handcrafted.
I think the "next big thing" with AI hardware will be when they switch from "digital" implementations of LLMs to "analogue". We already know that we can lose some bits of precision and still have a "workable" model. If/when folks figure the fine-tuning out, I'm guessing it'll be another order of magnitude improvement.
Token velocity is great, but the industry is hyper-fixated on speed while completely ignoring the blast radius. If we push to 17k tokens/sec for autonomous agents, we are just accelerating how fast an agent can hit an infinite loop and drain an API budget. Before we make AI ubiquitous, we need deterministic, network-level circuit breakers. Speed without governance is just a faster way to burn capital.
If it's not reprogrammable, it's just expensive glass.
If you etch the bits into silicon, you then have to accommodate the bits by physical area, which is the transistor density for whatever modern process they use. This will give you a lower bound for the size of the wafers.
This can give huge wafers for a very set model which is old by the time it is finalized.
Etching generic functions used in ML and common fused kernels would seem much more viable as they could be used as building blocks.
Models don’t get old as fast as they used to. A lot of the improvements seem to go into making the models more efficient, or the infrastructure around the models. If newer models mainly compete on efficiency it means you can run older models for longer on more efficient hardware while staying competitive.
If power costs are significantly lower, they can pay for themselves by the time they are outdated. It also means you can run more instances of a model in one datacenter, and that seems to be a big challenge these days: simply building an enough data centres and getting power to them. (See the ridiculous plans for building data centres in space)
A huge part of the cost with making chips is the masks. The transistor masks are expensive. Metal masks less so.
I figure they will eventually freeze the transistor layer and use metal masks to reconfigure the chips when the new models come out. That should further lower costs.
I don’t really know if this makes sanse. Depends on whether we get new breakthroughs in LLM architecture or not. It’s a gamble essentially. But honestly, so is buying nvidia blackwell chips for inference. I could see them getting uneconomical very quickly if any of the alternative inference optimised hardware pans out
From my own experience, models are at the tipping point for being useful at prototypes in software, and those are very large frontier models not feasible to get down on wafers unless someone does something smart.
I really don't like the hallucination rate for most models but it is improving, so that is still far in the future.
What I could see though, is if the whole unit they made would be power efficient enough to run on a robotics platform for human computer interaction.
It makes sense they would try to make repurposing their tech as much as they could since making changes is frought with a long time frame and risk.
But if we look long term and pretend that they get it to work, they just need to stay afloat until better smaller models can be made with their technology, so it becomes a waiting game for investors and a risk assessment.
> From my own experience, models are at the tipping point for being useful at prototypes in software
You must not have much experience using the new frontier models then. A lot of large tech companies are replacing their SDLC with agentic workflows. The tooling and frameworks are still ramping up, but the models have no problem producing production ready software given proper specifications.
Reading the in depth article also linked in this thread, they say that only 2 layers need to change most of the time. They claim from new model to PCB in 2 months. Let's see, but sounds promising.
I'm curious how much of "hardcoding" is in the chip? Can it have parts that don't need changing much and "offload" the rest into some sort of high-speed/bandwidth interconnect?
Will we reach a state where we have chips on which models can be "flashed" like CPU firmware?
Or eventually will we reach a state where none of these tricks will be needed because like run-of-the mill Intel/AMD commodity CPUs, we will have full-power AI chips which will be part of an bigger/integrated mother-chip? Then what will happen to companies that do LLMs-as-a-service? Will they be forced to join and adapt becoming hybrid model+hardware shops?
I'm not knowledgeable enough about hardware but throwing these random ideas out in hopes of thought-provoking responses to learn from.
These chips are large by fab standards and even with state of the art processes we likely won't see any kind of integration on consumer tech any time soon, but I imagine they will absolutely see instant demand if they can deliver on what they laid out in the post.
>Founded 2.5 years ago, Taalas developed a platform for transforming any AI model into custom silicon. From the moment a previously unseen model is received, it can be realized in hardware in only two months.
So this is very cool. Though I'm not sure how the economics work out? 2 months is a long time in the model space. Although for many tasks, the models are now "good enough", especially when you put them in a "keep trying until it works" loop and run them at high inference speed.
Seems like a chip would only be good for a few months though, they'd have to be upgrading them on a regular basis.
Unless model growth plateaus, or we exceed "good enough" for the relevant tasks, or both. The latter part seems quite likely, at least for certain types of work.
On that note I've shifted my focus from "best model" to "fastest/cheapest model that can do the job". For example testing Gemini Flash against Gemini Pro for simple tasks, they both complete the task fine, but Flash does it 3x cheaper and 3x faster. (Also had good results with Grok Fast in that category of bite-sized "realtime" workflows.)
1. Generic, mask layers and board to handle what's common across models. Especially memory and interface.
2. Specific layers for the model implementation.
Masks are the most expensive part of ASIC design. So, keeping the custom part small with the rest pre-proven in silicon, even shared across companies, would drop the costs significantly. This is already done in hardware industry in many ways but not model acceleration.
Then, do 8B, 30-40B, 70B, and 405B models in hardware. Make sure they're RLHF-tuned well since changes will be impossible or limited. Prompts will drive most useful functionality. Keep cranking out chips. There's maybe a chance to keep the weights changeable on-chip but it should still be useful if only inputs can change.
The other concept is to use analog, neural networks with the analog layers on older, cheaper nodes. We only have to customize that per model. The rest is pre-built digital with standard interfaces on a modern node. Given the chips would be distributed, one might get away with 28nm for the shared part and develop it eith shuttle runs.
I always thought eventually someone would come along and make a hardware accelerator for LLMs, but I thought it would be like google TPUs where you can load up whatever model you want. Baking the model into hardware sounds like the monkey paw curled, but it might be interesting selling an old.. MPU..? because it wasn't smart enough for your latest project
This is an interesting piece of hardware though when they go multi-chip for larger models the speed will no doubt suffer.
They'll also be severely limited on context length as it needs to sit in SRAM. Looks like the current one tops out at 6144 tokens which I presume is a whole chips worth. You'd also have to dedicate a chip to a whole user as there's likely only enough SRAM for one user's worth of context. I wonder how much time it takes them to swap users in/out? I wouldn't be surprised if this chip is severely underutilized (can't use it all when running decode as you have to run token by token with one users and then idle time as you swap users in/out).
Maybe a more realistic deployment would have chips for linear layers and chips for attention? You could batch users through the shared weight chips and then provision more or less attention chips as you want which would be per user (or shared amongst a small group 2-4 users).
To the authors: do not self-deprecate your work. It is true this is not a frontier model (anymore) but the tech you've built is truly impressive. Very few hardware startups have a v1 as good as this one!
Also, for many tasks I can think of, you don't really need the best of the best of the best, cheap and instant inference is a major selling point in itself.
I am thinking if this can be a low-level substrate for composing dumb LLMs into smart swarm, theoretically:
1. A whole with disparate parts (smart and dumb components) are almost always more cost-effective to reach a given target of performance
2. With that, a whole with disparate parts, are almost always more performant with the same cost
A few inspiration:
1. Human body is intelligent composed of so diverse parts
2. Swarm intelligence of insects and small animals are certainly beyond current understanding
The cost and speed of this thing is on point to make such a whole composed diverse parts possible.
The model routing discussion is fascinating. We're seeing similar patterns in how startups approach global talent - the best solution depends heavily on context. For engineering teams distributed across timezones (we work a lot with LATAM developers), the real bottleneck isn't just the model or the tool, it's understanding when to apply what. Same with inference: a 16k tok/s chip is incredible for real-time voice agents, but most startup use cases don't need that latency. The interesting question is whether we'll see more specialized hardware for niche applications, or if general-purpose solutions will keep winning through sheer volume economics.
I think the thing that makes 8b sized models interesting is the ability to train unique custom domain knowledge intelligence and this is the opposite of that. Like if you could deploy any 8b sized model on it and be this fast that would be super interesting, but being stuck with llama3 8b isn't that interesting.
The "small model with unique custom domain knowledge" approach has a very low capability ceiling.
Model intelligence is, in many ways, a function of model size. A small model tuned for a given domain is still crippled by being small.
Some things don't benefit from general intelligence much. Sometimes a dumb narrow specialist really is all you need for your tasks. But building that small specialized model isn't easy or cheap.
Engineering isn't free, models tend to grow obsolete as the price/capability frontier advances, and AI specialists are less of a commodity than AI inference is. I'm inclined to bet against approaches like this on a principle.
> Engineering isn't free, models tend to grow obsolete as the price/capability frontier advances, and AI specialists are less of a commodity than AI inference is. I'm inclined to bet against approaches like this on a principle.
This does not sound like it will simplify the training and data side, unless their or subsequent models can somehow be efficiently utilized for that.
However, this development may lead to (open source) hardware and distributed system compilation, EDA tooling, bus system design, etc getting more deserved attention and funding.
In turn, new hardware may lead to more training and data competition instead of the current NVIDIA model training monopoly market.
So I think you're correct for ~5 years.
A fine tuned 1.7B model probably is still too crippled to do anything useful. But around 8b the capabilities really start to change. I’m also extremely unemployed right now so I can provide the engineering.
Asking it what its knowledge cut-off is interesting, it doesn't seem to be consistent even within a single response. Sometimes it responds to say it cuts off 2020 too.
You
What is your knowledge cut-off?
Jimmy
My knowledge cut-off is 2022, which means that my training data is current up to 2021, but
I have been trained on a dataset that is updated periodically. If you have any specific
questions about events or topics that occurred after 2021, I may not have information on
those topics. However, I can still provide general information and context on those topics
to help guide further research.
The instantaneous response is impressive though. I'm sure there will be applications for this, I just lack the imagination to know what they'll be.
Sounds like they aren’t providing a system prompt which tells it what its knowledge cutoff is. It doesn’t know unless it’s told. Given the limited context, it makes sense they wouldn’t use up valuable space for that.
The system prompt is configurable in the sidebar and minimal. It doesn't give a knowledge cutoff. This is a tech demo of the fact it works at all, it's not meant to be a good chatbot.
So they create a new chip for every model they want to support, is that right? Looking at that from 2026, when new large models are coming out every week, that seems troubling, but that's also a surface take. As many people here know better than I that a lot of the new models the big guys release are just incremental changes with little optimization going into how they're used, maybe there's plenty of room for a model-as-hardware model.
Which brings me to my second thing. We mostly pitch the AI wars as OpenAI vs Meta vs Claude vs Google vs etc. But another take is the war between open, locally run models and SaaS models, which really is about the war for general computing. Maybe a business model like this is a great tool to help keep general computing in the fight.
We’re reaching a saturation threshold where older models are good enough for many tasks, certainly at 100x faster inference speeds. Llama3.1 8B might be a little too old to be directly useful for e.g. coding but it certainly gets the gears turning about what you could do with one Opus orchestrator and a few of these blazing fast minions to spit out boilerplate…
One of these things, however old, coupled with robust tool calling is a chip that could remain useful for decades. Baking in incremental updates of world knowledge isn't all that useful. It's kinda horrifying if you think about it, this chip among other things contains knowledge of Donald Trump encoded in silicon. I think this is a way cooler legacy for Melania than the movie haha.
This is incredible. With this speed I can use LLMs in a lot of pre-filtering etc. tasks. As a trivial example, I have a personal OpenClaw-like bot that I use to do a bunch of things. Some of the things just require it to do trivial tool-calling and tell me what's up. Things like skill or tool pre-filtering become a lot more feasible if they're always done.
Anyway, I imagine these are incredibly expensive, but if they ever sell them with Linux drivers and slotting into a standard PCIe it would be absolutely sick. At 3 kW that seems unlikely, but for that kind of speed I bet I could find space in my cabinet and just rip it. I just can't justify $300k, you know.
The demo was so fast it highlighted a UX component of LLMs I hadn’t considered before: there’s such a thing as too fast, at least in the chatbot context. The demo answered with a page of text so fast I had to scroll up every time to see where it started. It completely broke the illusion of conversation where I can usually interrupt if we’re headed in the wrong direction. At least in some contexts, it may become useful to artificially slow down the delivery of output or somehow tune it to the reader’s speed based on how quickly they reply. TTS probably does this naturally, but for text based interactions, still a thing to think about.
That is what self-driving car should eventually use, whenever they (or the authorities) deem their model good enough. Burn it on a dedicated chip. It would be cheaper (energy) to run, and faster to make decisions.
> It seems like "six seven" is likely being used to represent the number 17. Is that correct? If so, I'd be happy to discuss the significance or meaning of the number 17 with you.
This is genuinely an incredible proof-of-concept; the business implications of this demo to the AI labs and all the companies that derive a ton of profit from inference is difficult to understate, really.
I think this is how I'm going to get my dream of Opus 3.7 running locally, quickly and cheaply on my mid-tier MacBook in 2030. Amazing. Anthropic et al will be able to make marginal revenue from licensing the weights of their frontier-minus-minus models to these folks.
I do like the idea of an aftermarket of ancient LLM chips that still have tons of useful life on text processing tasks etc. They don't talk about their architecture much, I wonder how well power can scale down. 200W for such a small model is not something I see happening in a laptop any time soon. Pretty hilarious implications for moat-building of the big providers too.
Yea I mean this is the first publishable draft of a startup cooking on this.
I'm confident there are at least 1-2 OOMs of improvement to come here in terms of the (intelligence : wattage) ratio.
I really thought we were going to need to see a couple of dramatic OOM-improvement changes to the model composition / software layer, in order to get models of Opus 3.7's capability running on our laptops.
This release tells me that eventual breakthrough won't even be strictly necessary, imo.
The way I imagine it in 2-4 years we're going to be hit with a triple glut of better architecture, massive oversupply of hardware and potentially one or two hardware efforts like this really taking off. It's pretty crazy we're already 4 years in and outside of very niche / low availability solutions, it's still either GPU or bust
That's interesting! How do you see "oversupply of hardware" playing out?
Is it because we stop doing ~2024-style, large-scale training (marginal returns aren't worth it)? Or because supply way outpaces the training+inference demand?
AFAIU if the trend lines /S-curves keep chugging along as they are, we won't hit hardware oversupply for a long, long time without some sort of AI training winter.
I think frontier models can do more with fewer tokens (and do the wrong thing far less often) than a "really fast" small model.
There are use cases for fast/ultrafast inferrence models - classifying text, scoring things, extracting information - but for coding and other knowledge tasks - you're not going to get to your solution faster at 16,000 tokens/s if the solution never comes (or is the wrong one).
I think this is quite interesting for local AI applications. As this technology basically scales with parameter size, if there could be some ASIC for a QWen 0.5B or Google 0.3B model thrown onto a laptop motherboard it'd be very interesting.
Obviously not for any hard applications, but for significantly better autocorrect, local next word predictions, file indexing (tagging I suppose).
The efficiency of such a small model should theoretically be great!
Are the model weights burned into the silicon / part of the architecture? Or can you update the model weights on these chips?
If they cannot be updated, these chips will be outdated the moment they are made given the breakneck speed at which new and improved models are introduced.
The implications for RLM is really interesting. RLM is expensive because of token economics. But when tokens are so cheap and fast to generate, context size of the model matters a lot less
Also interesting implications for optimization-driven frameworks like DSPy. If you have an eval loop and useful reward function, you can iterate to the best possible response every time and ignore the cost of each attempt
17k token/sec is $0.18/chip/hr for the size of H100 chip if they want to compete with the market rate[1]. But 17k token/sec could lead to some new usecases.
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Insane gains, makes me excited for the future. Imagine Opus-like responses in <1 second.
I suspect power efficiency will be nearly entirely offset by increased usage but it’s more bang for watt.
Embedding the model at chip fab time ought to be useful for robotics, driving, vision, and audio applications, at least. The training sets are good for years.
So they use 3 bit values. Is that current thinking? LLMs started at 32-bit floats, and have gradually shrunk. 8-bit floats seem to work. Is 3 bits pushing it?
This is really cool! I am trying to find a way to accelerate LLM inference for PII detection purposes, where speed is really necessary as we want to process millions of log lines per minute, I am wondering how fast we could get e.g. llama 3.1 to run on a conventional NVIDIA card? 10k tokens per second would be fantastic but even at 1k this would be very useful.
Yeah I mean we have a mechanism that can bypass AI models for log lines where we are pretty sure no PII is in there (kind of like smart caching using fuzzy template matching to identify things that we have seen before many times, as logs tend to contain the same stuff over and over with tiny variations e.g. different timestamps), so we only need to pass the lines where we cannot be sure there's nothing to the AI for inspection. And we can of course parallelize. Currently we use a homebrew CFR model with lots of tweaks and it's quite good but an LLM would of course be much better still and capture a lof of cases that would evade the simpler model.
For that you only need high throughput which is much easier to achieve compared to high latency, thanks to batching -- assuming the log lines or chunks can be processed independently. You can check TensorRT-LLM benchmarks (https://nvidia.github.io/TensorRT-LLM/developer-guide/perf-o...), or try running vllm on a card you have access to.
Maybe its the tism but I also read that sentence as neutral. You expected very little and you got very little. Why would that be positive or negative? Maybe it should be positive because you got what you were expecting? But I would call getting what you expect something neutral, if you expected little and got a lot then that would be positive. If you expected a lot and got little then its negative. But if you expected little and got little the most clear outcome is that its a neutral statement. Am I missing something?
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Am I reading this right: 10x faster and 10x less power, ie. 100x more power efficient?
There are a lot of problems solved by tiny models. The huge ones are fun for large programming tasks, exploration, analysis, etc. but there's a massive amount of processing <10GB happening every day. Including on portable devices.
This is great even if it can't ever run Opus. Many people will be extremely happy about something like Phi accessible at lightning speed.
Positive results in mice also known as being a promising proof of concept. At this point, anything which deflates the enormous bubble around GPUs, memory, etc, is a welcome remedy. A decent amount of efficient, "good enough" AI will change the market very considerably, adding a segment for people who don't need frontier models. I'd be surprised if they didn't end up releasing something a lot bigger than they have.
This is crazy! These chips could make high-reasoning models run so fast that they could generate lots of solution variants and automatically choose the best. Or you could have a smart chip in your home lab and run local models - fast, without needing a lot of expensive hardware or electricity
If they made a low power/mobile version, this could be really huge for embedded electronics. Mass produced, highly efficient "good enough" but still sort of dumb ais could put intelligence in house hold devices like toasters, light switches, and toilets. Truly we could be entering into the golden age of curses.
Oh god, this is the new version of every device having Bluetooth and an app and being called “smart”.
I just wanted some toast, but here I am installing an app, dismissing 10 popups, and maybe now arguing with a chat bot about how I don’t in fact want to turn on notifications.
I totally buy the thesis on specialization here, I think it makes total sense.
Asides from the obvious concern that this is a tiny 8B model, I'm also a bit skeptical of the power draw. 2.4 kW feels a little bit high, but someone else should try doing the napkin math compared to the total throughput to power ratio on the H200 and other chips.
This makes me think about how large would an FPGA-based system to be able to do this? Obviously there is no single-chip FPGA that can do this kind of job, but I wonder how many we would need.
Also, what if Cerebras decided to make a wafer-sized FPGA array and turned large language models into lots and lots of logical gates?
This is pretty wild! Only Llama3.1-8B, but this is only their first release so you can assume they're working on larger versions.
So what's the use case for an extremely fast small model? Structuring vast amounts of unstructured data, maybe? Put it in a little service droid so it doesn't need the cloud?
I imagine how advantageous it would be to have something like llama.cpp encoded on a chip instead, allowing us to run more than a single model. It would be slower than Jimmy, for sure, but depending on the speed, it could be an acceptable trade-off.
I am super happy to see people working on hardware for local llm. Yet, isnt it premature ? Space is still evolving. Today, i refuse to buy a gpu because i do not know what will be the best model tomorrow.
Waiting to get a on the shelf device to run an opus like model
Strange that they apparently raised $169M (really?) and the website looks like this. Don't get me wrong: Plain HTML would do if "perfect", or you would expect something heavily designed. But script-kiddie vibe coded seems off.
Strange that they raised money at all with an idea like this.
It's a bad idea that can't work well. Not while the field is advancing the way it is.
Manufacturing silicon is a long pipeline - and in the world of AI, one year of capability gap isn't something you can afford. You build a SOTA model into your chips, and by the time you get those chips, it's outperformed at its tasks by open weights models half their size.
Now, if AI advances somehow ground to a screeching halt, with model upgrades coming out every 4 years, not every 4 months? Maybe it'll be viable. As is, it's a waste of silicon.
The prototype is: silicon with a Llama 3.1 8B etched into it. Today's 4B models already outperform it.
Token rate in five digits is a major technical flex, but, does anyone really need to run a very dumb model at this speed?
The only things that come to mind that could reap a benefit are: asymmetric exotics like VLA action policies and voice stages for V2V models. Both of which are "small fast low latency model backed by a large smart model", and both depend on model to model comms, which this doesn't demonstrate.
In a way, it's an I/O accelerator rather than an inference engine. At best.
Even if this first generation is not useful, the learning and architecture decisions in this generation will be. You really can't think of any value to having a chip which can run LLMs at high speed and locally for 1/10 of the energy budget and (presumably) significantly lower cost than a GPU?
If you look at any development in computing, ASICs are the next step. It seems almost inevitable. Yes, it will always trail behind state of the art. But value will come quickly in a few generations.
maybe they're betting on improvement in models to plateau, and that having a fairly stablized capable model that is orders of magnitude faster than running on GPU's can be valuable in the future?
the non-frontier model use case is underrated here. for things like autocomplete, classification, routing, or first-pass triage -- you don't need GPT-4 class reasoning. you need something that's fast and cheap enough to call on every keystroke or every incoming message. at 17k tok/sec and 20x lower cost, a whole class of latency-sensitive edge applications becomes viable that aren't today.
But as models are changing rapidly and new architectures coming up, how do they scale and also we do t yet know the current transformer architecture will scale more than it already is. Soo many ope questions but VCs seems to be pouring money.
It would be pretty incredible if they could host an embedding model on this same hardware, I would pay for that immediately. It would change the type of things you could build by enabling on the fly embeddings with negligible latency.
Imagine a mass-produced AI chips with all human knowledge packed in chinesium epoxy blobs running from CR2032 batteries in toys for children. Given the progress in density and power consumption, it's not that far away.
The token throughput improvements are impressive. This has direct implications for usage-based billing in AI products — faster inference means lower cost per request, which changes the economics of credits-based pricing models significantly.
So I'm guessing this is some kind of weights as ROM type of thing? At least that's how I interpret the product page, or maybe even a sort of ROM type thing that you can only access by doing matrix multiplies.
You shouldn't need any ROM. It's likely the architecture is just fixed hardware with weights loaded in via scan flip-flows. If it was me making it, I'd just design a systolic array. Just multipliers feeding into multipliers, without even going through RAM.
That seems promising for applications that require raw speed. Wonder how much they can scale it up - 8B model quantized is very usable but still quite small compared to even bottom end cloud models.
Would it make sense for the big players to buy them? Seems to be a huge avenue here to kill inference costs which always made me dubious on LLMs in general.
Performance like that may open the door to the strategy of brutefocing solutions to problems for which you have a verifier (problems such as decompilation).
For fun I'm imagining a future where you would be able to buy an ASIC with like an hard-wired 1B LLM model in it for cents and it could be used everywhere.
Yep, this is the most exciting demo for me yet. Holy cow this is unbelievably fast.
The most impressive demo since gpt 3, honestly.
Since we already have open source models that are plenty good, like the new kimi k2.5, all I need is the ability to run it at moderate speed.
Honestly I am not bullish on capabilities that models do not yet have, seems we have seen it all and the only advancement have been context size.
And honestly I would claim this is the market sentiment aswell, anthropic showed opus 4.6 first and the big release was actually sonnet, the model people would use routinely.
Nobody gave a shit about Gemini 3.1 pro, 3.0 flash was very successful...
Given all the recent developments in the last 12 months, no new use cases have opened for me.
Given this insane speed, even on a limited model/context size, we would approach IA very differently.
There is also the use case of delegating tasks programmatically to an LLM, for example, transforming unstructured data to structured data. This task often can’t be done reliably without either 1. lots of manual work, or 2. intelligence, especially when the structure of the individual data pieces are unknown. Problems like these can be much more efficiently solved by LLMs, and if you imagine these programs are processing very large datasets, then sub-millisecond inference is crucial.
Agents also "read", so yes there is. Think about spinning up 10, 20, 100 sub agents for a small task and they all return near instant. That's the usecase, not the chatbot.
I know it is not easy to see the benefits of small models easily but this is what I am building for (1). I created a product for Google Gemini 3 Hackathon and I used Gemini 3 Flash (2). I tested locally using Ministral 3B and it was promising. Definitely will need work. But 8B/14B may give awesome results.
I am building a data extraction software on top of emails, attachments, cloud/local files. I use a reverse template generation with only variable translation done by LLMs (3). Small models are awesome for this (4).
I just applied for API access. If privacy policies are a fit, I would love to enable this for MVP launch.
I still believe this is the right - and inevitable - path for AI, especially as I use more premium AI tooling and evaluate its utility (I’m still a societal doomer on it, but even I gotta admit its coding abilities are incredible to behold, albeit lacking in quality).
Everyone in Capital wants the perpetual rent-extraction model of API calls and subscription fees, which makes sense given how well it worked in the SaaS boom. However, as Taalas points out, new innovations often scale in consumption closer to the point of service rather than monopolized centers, and I expect AI to be no different. When it’s being used sparsely for odd prompts or agentically to produce larger outputs, having local (or near-local) inferencing is the inevitable end goal: if a model like Qwen or Llama can output something similar to Opus or Codex running on an affordable accelerator at home or in the office server, then why bother with the subscription fees or API bills? That compounds when technical folks (hi!) point out that any process done agentically can instead just be output as software for infinite repetition in lieu of subscriptions and maintained indefinitely by existing technical talent and the same accelerator you bought with CapEx, rather than a fleet of pricey AI seats with OpEx.
The big push seems to be building processes dependent upon recurring revenue streams, but I’m gradually seeing more and more folks work the slop machines for the output they want and then put it away or cancel their sub. I think Taalas - conceptually, anyway - is on to something.
> Though society seems poised to build a dystopian future defined by data centers and adjacent power plants, history hints at a different direction. Past technological revolutions often started with grotesque prototypes, only to be eclipsed by breakthroughs yielding more practical outcomes.
…for a privileged minority, yes, and to the detriment of billions of people whose names the history books conveniently forget. AI, like past technological revolutions, is a force multiplier for both productivity and exploitation.
Maybe not today. Opus is quite large. This demo works with a very small 8B model. But, maybe one day. Hopefully soon. Opus on a chip would be very awesome, even if it can never be upgraded.
Someone mentioned that maybe we'd see a future where these things come in something like Nintendo cartridges. Want a newer model? Pop in the right catridge.
I have a hard time reading beyond factual lies like:
> On the cost front, deploying modern models demands massive engineering and capital: room-sized supercomputers consuming hundreds of kilowatts…
This is just wrong. The largest models are probably 1-2 trillion parameters. Say 2 trillion and let’s pretend it’s only quantized to 8bit (even though it could easily be half that.) So we need 2TB of VRAM. Not even using the latest hardware, lets say H100 chips with 80GB of vram each, with 8 of them in say an 8U. (Although you can certainly fit these in 6U still air cooled or even 4U water cooled.) Three of these server would almost do, but let’s call it four to include plenty of room for context. The largest physical size would be 32U - most of a single rack. Which is hardly the size of a room, even in Manhattan. Total power maybe 40kW. And you could easily drop these numbers to a half or quarter of that with reasonable modifications or upgrades.
If you want to sell your hardware, start by being honest about the problem you’re addressing.
"Many believe AI is the real deal. In narrow domains, it already surpasses human performance. Used well, it is an unprecedented amplifier of human ingenuity and productivity."
Sounds like people drinking the Kool-Aid now.
I don't reject that AI has use cases. But I do reject that it is promoted as "unprecedented amplifier" of human xyz anything. These folks would even claim how AI improves human creativity. Well, has this been the case?
Same, I would have given up on them long ago, I no longer code at all now. Why would I when the latest models can do it better, faster and without the human limitations of tiredness, emotional impacts etc.
> These folks would even claim how AI improves human creativity. Well, has this been the case?
Yes. Example: If you've never programmed in language X, but want to build something in it, you can focus on getting from 0 to 1 instead of being bogged down in the idiosyncrasies of said language.
There's a scifi story here when millions of these chips, with Qwen8-AGI-Thinking baked into them, are obsoleted by the release of Qwen9-ASI, which promptly destroys humanity and then itself by accident. A few thousand years later, some of the Qwen8 chips in landfill somehow power back up again and rebuild civilization on Earth.
This is, so far, utterly charming. I made a simple prompt of "make an adventure game in the style of cia.bas from pc-sig". It ended up being wildly different than that, but 30 minutes later and I'm still busy trying to play this "game" it fabricated out of thin air.
One interesting thing is it keeps randomly emitting "ประก" (meaning "Announcement") and chartInstance. This is recalling the early days of GPT-2 when the light bulb went on that "hey, there's something groundbreaking here".
Whoever doesn’t buy/replicate this in the next year is dead. Imagine OpenAI trying to sell you a platform that takes 15 minutes, when someone else can do it in 0.001s.
> Generate a fiction book about an old sorcerer who, like all elderly magicians, has to go back to studying magic for two years at a magic school in order to be granted the right to keep his wand and to keep doing magic at his home.
Answer, generated in 0,063s @ 15 742 tok/s:
> The Relearning of Eryndor Thorne
> In the quaint village of Brindlemark, where ancient traditions and mystical energies coexisted, Eryndor Thorne had lived a life of quiet reverence. For nigh on five decades, he had wielded the arcane arts with precision and mastery, earning him the respect of his peers and the awe of the villagers. His home, a humble abode atop a hill overlooking the rolling hills and verdant valleys, was a sanctuary of sorts, a place where the very fabric of magic was woven into the air. His library was a treasure trove of ancient tomes, texts, and scrolls, filled with the wisdom of the ages. Yet, despite his mastery over the arcane, Eryndor felt the stirrings of a nagging concern. The once-renewable license to practice magic, a privilege granted by the Council of Mages, was on the cusp of expiration.
> As all elderly mages knew, every two decades, the Council would demand that those who had spent a lifetime mastering the arcane arts return to the prestigious Wysteria Academy, the hallowed halls of magic learning, to relearn the foundational principles of magic. This was a ritual as old as the craft itself, a test of the elderly mage's understanding, to prove they still possessed the fundamental knowledge and skills required to practice magic with precision and safety.
> Eryndor, like many of his fellow mages, had grown complacent in his mastery. He had mastered the intricacies of elemental magic, bending the winds, waters, and flames to his will. His wand, a family heirloom passed down through generations, had become an extension of his arm, channeling the power of magic with ease. Yet, despite his proficiency, the Council's decree was clear: Eryndor must return to Wysteria Academy for two years of reeducation.
> The journey was not an easy one. Packing his worn leather satchel with the few books and scrolls he deemed essential, Eryndor bid farewell to his loved ones in Brindlemark, leaving behind the familiarity of his home to embark on a journey that would take him to a place both wondrous and treacherous. At Wysteria Academy, he joined a cohort of students from all corners of the realm, each as skilled and eager as he once was. The relearning process was a grueling one, as if facing a mountain he had once conquered. New techniques were presented, new theories to unravel, and new expectations to meet. Eryndor, once a master, now sat at the feet of his teachers, absorbing the new knowledge like a parched traveler at an oasis.
> Among his fellow students, there was the young, fire-kissed mage, Elara, who wielded magic with an intensity that bordered on reckless abandon. Her fiery nature and quick wit often put her at odds with the strict, ancient traditions, earning her a certain notoriety among the academy's elder mages. Then there was the enigmatic, shadow-drawn Kael, whose mastery of the arcane was matched only by his mystery. Kael's affinity for the dark arts raised more than a few eyebrows among the faculty, but Eryndor, having once walked the fine line between light and shadow, saw something of himself in the young mage.
> As the years passed, Eryndor grew to appreciate the challenges and opportunities that came with his return to the academy. He found himself grappling with the nuances of magic anew, rekindling memories of his early days as a novice. The relearning process was as much about rediscovering himself as it was about mastering the arcane. His studies were a journey of self-discovery, one that tested the mettle of his will and the depths of his understanding.
> Upon completion of his studies, Eryndor stood before the Council once more, his wand in hand, his heart afire with the thrill of rediscovery. The Council's examination was not merely a test of his knowledge but a test of his character. Eryndor, like many of his peers, had grown complacent, but the rigors of relearning had rekindled a spark within him, a flame that would guide him through the trials ahead.
> With his renewed license granted, Eryndor returned to Brindlemark, his home and his heart rejuvenated. His library, once a testament to his mastery, was now a canvas for the new knowledge he had acquired. His wand, now an extension of his rekindled passion for the arcane, channelled magic with a precision and flair that was unmatched. Eryndor Thorne, the elderly mage, had not only relearned magic but had rediscovered himself, a testament to the transformative power of education and the enduring spirit of the arcane.
I don't know why, but my ultra wide monitor absolutely hates that site. The whole screen is flickering trying to deal with the annoying background. Thank the gods for reader mode.
he numbers being presented are deliberately misleading. On this model, Groq delivers around 1,300 tokens per second, whereas Cerebras achieves roughly 2,500 tokens per second.
With the next generation of Cerebras chips expected to be 5–7× faster, peak throughput could reach the ~17,500 tokens-per-second range. For smaller models like this, that level of performance is entirely realistic. So no, a general-purpose accelerator will likely continue to outperform a fixed-function ASIC with a specific model etched into it.
Moreover, we’re only looking at results from a two-year-old, relatively small model. We still don’t know how this architecture will scale with a large MoE model, especially given constraints like limited on-chip KV cache and more complex attention mechanisms.
The real test isn’t performance on a small benchmark model, it’s how the system handles large-scale, production-grade workloads under architectural constraints.
User
You: What is the airspeed of an unladen swallow/
Jimmy:
A reference to the classic line from Monty Python and the Holy Grail!
Thus far it is correct but here it comes:
In the movie, a group of medieval scholars, the Knights who say "Ni!" (also known as the Bridgekeepers), ask a king's servant, a blacksmith, if an unladen swallow makes good time. The blacksmith, played by John Cleese, responds with the now-famous line:
"And what's the airspeed velocity of an unladen swallow?"
To which the blacksmith responds:
"What do you mean? An unladen swallow?"
The blacksmith then explains that an unladen swallow is when a swallow (a type of bird) is flying in the air without carrying any burden, implying that it is moving at a speed that is not affected by carrying any weight or load.
Nope, that is not at all how it went. No blacksmiths, not related to the Knights who say 'Ni!' - who have nothing scholarly about them, they wanted a shrubbery and demanded another one when their wish was fulfilled. It was the Bridgekeeper - not related to the aforementioned knights - who asked 3 questions and would throw anyone who could not answer all three into the chasm. When he asked Arthur - played by Cleese - the question about the airspeed of the swallow he asked the Bridgekeeper whether he meant an African or European swallow. The Bridgekeeper did not have an answer and was thrown into the chasm, problem solved.
However, in reality, swallows are not typically "unladen" in flight. They are small birds that fly at relatively fast speeds, usually around 10-15 km/h (6-9 mph), but they are not usually carrying any burdens!
Needless LLM-blabber.
The "airspeed velocity of an unladen swallow" has become a meme and a cultural reference point, often used humorously or ironically to refer to situations where someone is trying to make an absurd or non-sensical argument or ask an absurd question.
Somewhat correct but not necessary in this context.
The correct answer to the question would have been Do you mean an African or European swallow? followed by a short reference to the movie.
Of course this demo is not about the accuracy of the model - 'an old Llama' as mentioned elsewhere in this thread - but it does show that speed isn't everything. For generating LLM-slop this hardware implementation probably offers an unbeatable price/performance ratio but it remains to be seen if it can be combined with larger and less hallucination-prone models.
Is it already available to buy, or is this a “pay now, get it later” kind of new ASIC miner?
Sorry for being skeptical, but AI is the new "crypto coin", and the crypto bros are still around.
There doesn't seem to be any form of buying the HC1 hardware at the moment. There is a free chatbot demo and then a form to request access to API. They seem to intend HC1 to be for demonstration and HC2 for "real" use, but they don't seem to be taking payment for either at the moment.
Scale this then close the loop and have fabs spit out new chips with latest weights every week that get placed in a server using a robot, how long before AGI?
The article doesn't say anything about the price (it will be expensive), but it doesn't look like something that the average developer would purchase.
An LLM's effective lifespan is a few months (ie the amount of time it is considered top-tier), it wouldn't make sense for a user to purchase something that would be superseded in a couple of months.
An LLM hosting service however, where it would operate 24/7, would be able to make up for the investment.
This is not a general purpose chip but specialized for high speed, low latency inference with small context. But it is potentially a lot cheaper than Nvidia for those purposes.
Tech summary:
This is all from their website, I am not affiliated. The founders have 25 years of career across AMD, Nvidia and others, $200M VC so far.Certainly interesting for very low latency applications which need < 10k tokens context. If they deliver in spring, they will likely be flooded with VC money.
Not exactly a competitor for Nvidia but probably for 5-10% of the market.
Back of napkin, the cost for 1mm^2 of 6nm wafer is ~$0.20. So 1B parameters need about $20 of die. The larger the die size, the lower the yield. Supposedly the inference speed remains almost the same with larger models.
Interview with the founders: https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...
This math is useful. Lots of folks scoffing in the comments below. I have a couple reactions, after chatting with it:
1) 16k tokens / second is really stunningly fast. There’s an old saying about any factor of 10 being a new science / new product category, etc. This is a new product category in my mind, or it could be. It would be incredibly useful for voice agent applications, realtime loops, realtime video generation, .. etc.
2) https://nvidia.github.io/TensorRT-LLM/blogs/H200launch.html Has H200 doing 12k tokens/second on llama 2 12b fb8. Knowing these architectures that’s likely a 100+ ish batched run, meaning time to first token is almost certainly slower than taalas. Probably much slower, since Taalas is like milliseconds.
3) Jensen has these pareto curve graphs — for a certain amount of energy and a certain chip architecture, choose your point on the curve to trade off throughput vs latency. My quick math is that these probably do not shift the curve. The 6nm process vs 4nm process is likely 30-40% bigger, draws that much more power, etc; if we look at the numbers they give and extrapolate to an fp8 model (slower), smaller geometry (30% faster and lower power) and compare 16k tokens/second for taalas to 12k tokens/s for an h200, these chips are in the same ballpark curve.
However, I don’t think the H200 can reach into this part of the curve, and that does make these somewhat interesting. In fact even if you had a full datacenter of H200s already running your model, you’d probably buy a bunch of these to do speculative decoding - it’s an amazing use case for them; speculative decoding relies on smaller distillations or quants to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model.
Upshot - I think these will sell, even on 6nm process, and the first thing I’d sell them to do is speculative decoding for bread and butter frontier models. The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
I hope these guys make it! I bet the v3 of these chips will be serving some bread and butter API requests, which will be awesome.
> any factor of 10 being a new science / new product category,
I often remind people two orders of quantitative change is a qualitative change.
> The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
The real product they have is automation. They figured out a way to compile a large model into a circuit. That's, in itself, pretty impressive. If they can do this, they can also compile models to an HDL and deploy them to large FPGA simulators for quick validation. If we see models maturing at a "good enough" state, even a longer turnaround between model release and silicon makes sense.
While I also see lots of these systems running standalone, I think they'll really shine combined with more flexible inference engines, running the unchanging parts of the model while the coupled inference engine deals with whatever is too new to have been baked into silicon.
I'm concerned with the environmental impact. Chip manufacture is not very clean and these chips will need to be swapped out and replaced at a cadence higher than we currently do with GPUs.
Having dabbled in VLSI in the early-2010s, half the battle is getting a manufacturing slot with TSMC. It’s a dark art with secret handshakes. This demonstrator chip is an enormous accomplishment.
Yeah and a team I’m not familiar with — I didn’t check bios but they don’t lead with ‘our team made this or that gpu for this or that bigco’.
The design ip at 6nm is still tough; I feel like this team must have at least one real genius and some incredibly good support at tsmc. Or they’ve been waiting a year for a slot :)
From the article:
"Ljubisa Bajic desiged video encoders for Teralogic and Oak Technology before moving over to AMD and rising through the engineering ranks to be the architect and senior manager of the company’s hybrid CPU-GPU chip designs for PCs and servers. Bajic did a one-year stint at Nvidia as s senior architect, bounced back to AMD as a director of integrated circuit design for two years, and then started Tenstorrent."
His wife (COO) worked at Altera, ATI, AMD and Testorrent.
"Drago Ignjatovic, who was a senior design engineer working on AMD APUs and GPUs and took over for Ljubisa Bajic as director of ASIC design when the latter left to start Tenstorrent. Nine months later, Ignjatovic joined Tenstorrent as its vice president of hardware engineering, and he started Taalas with the Bajices as the startup’s chief technology officer."
Not a youngster gang...
There might be a foodchain of lower order uses when they become "obsolete".
I think there will be a lot of space for sensorial models in robotics, as the laws of physics don't change much, and a light switch or automobile controls have remained stable and consistent over the last decades.
I think the next major innovation is going to be intelligent model routing. I've been exploring OpenClaw and OpenRouter, and there is a real lack of options to select the best model for the job and execute. The providers are trying to do that with their own models, but none of them offer everything to everyone at all times. I see a future with increasingly niche models being offered for all kinds of novel use cases. We need a way to fluidly apply the right model for the job.
Agree that routing is becoming the critical layer here. Vllm iris is really promising for this https://blog.vllm.ai/2026/01/05/vllm-sr-iris.html
There's already some good work on router benchmarking which is pretty interesting
At 16k tokens/s why bother routing? We're talking about multiple orders of magnitude faster and cheaper execution.
Abundance supports different strategies. One approach: Set a deadline for a response, send the turn to every AI that could possibly answer, and when the deadline arrives, cancel any request that hasn't yet completed. You know a priori which models have the highest quality in aggregate. Pick that one.
The best coding model won’t be the best roleplay one which won’t be the best at tool use. It depends what you want to do in order to pick the best model.
I'm not saying you're wrong, but why is this the case?
I'm out of the loop on training LLMs, but to me it's just pure data input. Are they choosing to include more code rather than, say fiction books?
I’ll go ahead and say they’re wrong (source: building and maintaining llm client with llama.cpp integrated & 40+ 3p models via http)
I desperately want there to be differentiation. Reality has shown over and over again it doesn’t matter. Even if you do same query across X models and then some form of consensus, the improvements on benchmarks are marginal and UX is worse (more time, more expensive, final answer is muddied and bound by the quality of the best model)
Thanks. Are there any links where I can learn more about this?
I did some Googling and it appears that there are some examples where people say combining multiple models or multiple runs of the same models leads to improvements: https://www.sciencedirect.com/science/article/abs/pii/S00104... https://arxiv.org/abs/2203.11171
But presumably people are less likely to publish a paper when an approach doesn’t work.
Are you saying I’m wrong that some models are better for some tasks than others, but there isn’t a universally best model for all tasks?
[dead]
There is the pre-training, where you passively read stuff from the web.
From there you go to RL training, where humans are grading model responses, or the AI is writing code to try to pass tests and learning how to get the tests to pass, etc. The RL phase is pretty important because it's not passive, and it can focus on the weaker areas of the model too, so you can actually train on a larger dataset than the sum of recorded human knowledge.
I came across this yesterday. Haven't tried it, but it looks interesting:
https://agent-relay.com/
[dead]
> speculative decoding for bread and butter frontier models. The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious
Can we use older (previous generation, smaller) models as a speculative decoder for the current model? I don't know whether the randomness in training (weight init, data ordering, etc) will affect this kind of use. To the extent that these models are learning the "true underlying token distribution" this should be possible, in principle. If that's the case, speculative decoding is an elegant vector to introduce this kind of tech, and the turnaround time is even less of a problem.
> The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.
They may be using Rapidus, which is a Japanese government backed foundry built around all single wafer processing vs traditional batching. They advertise ~2 month turnaround time as standard, and as short as 2 weeks for priority.
For speculative decoding, wouldn’t this be of limited use for frontier models that don’t have the same tokenizer as Llama 3.1? Or would it be so good that retokenization/bridging would be worth it?
My understanding as well is that speculative decoding only works with a smaller quant of the same model. You're using the faster sampling of the smaller models representation of the larger models weights in order to attempt to accurately predict its token output. This wouldn't work cross-model as the token probabilities are completely different.
This is not correct.
Families of model sizes work great for speculative decoding. Use the 1B with the 32B or whatever.
It's a balance as you want it to be guessing correctly as much as possible but also be as fast as possible. Validation takes time and every guess needs to be validated etc
The model you're using to speculate could be anything, but if it's not guessing what the main model would predict, it's useless.
> The model you're using to speculate could be anything, but if it's not guessing what the main model would predict, it's useless.
So what I said is correct then lol. If you're saying I can use a model that isn't just a smaller quant of the larger model I'm trying to speculatively decode, except that model would never get an accurate prediction, then how is that in any way useful or desirable?
Smaller quant or smaller model?
Afaik it can work with anything, but sharing vocab solves a lot of headaches and the better token probs match, the more efficient it gets.
Which is why it is usually done with same family models and most often NOT just different quantizations of the same model.
Smaller quant of the same model. A smaller quant of a different family of model would be practically useless and there wouldn't be any point in even setting it up.
I think they’d commission a quant directly. Benefits go down a lot when you leave model families.
Think about this for solving questions in math where you need to explore a search space. You can run 100 of these for the same cost and time of doing one api call to open ai.
The guts of a LLM isn't something I'm well versed in, but
> to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model
suggests there is something I'm unaware of. If you compare the small and big model, don't you have to wait for the big model anyway and then what's the point? I assume I'm missing some detail here, but what?
Speculative decoding takes advantage of the fact that it's faster to validate that a big model would have produced a particular sequence of tokens than to generate that sequence of tokens from scratch, because validation can take more advantage of parallel processing. So the process is generate with small model -> validate with big model -> then generate with big model only if validation fails
More info:
* https://research.google/blog/looking-back-at-speculative-dec...
* https://pytorch.org/blog/hitchhikers-guide-speculative-decod...
See also speculative cascades which is a nice read and furthered my understanding of how it all works
https://research.google/blog/speculative-cascades-a-hybrid-a...
Verification is faster than generation, one forward pass for verification of multiple tokens vs a pass for every new token in generation
I don't understand how it would work either, but it may be something similar to this: https://developers.openai.com/api/docs/guides/predicted-outp...
When you predict with the small model, the big model can verify as more of a batch and be more similar in speed to processing input tokens, if the predictions are good and it doesn't have to be redone.
They are referring to a thing called "speculative decoding" I think.
Most importantly this opens up an amazing future where we get the real version of the classic science fiction MacGuffin of a physical AI chip. Pair this with several TB of flash storage and you have persistent artificial consciousness that can be carried around with you. Bonus points if it's quirky, custom-trained and the chip is one of a kind that you stole from an evil corporation. Additional bonus points if the packaging is such that it's small enough to plug into the USB-C port on your smart glasses and has an eBPF module it can leverage to see what you're doing and talk to you in real time about your actions.
I enjoy envisioning futures more whimsical than "the bargain-basement LLM provider that my insurance company uses denied my claim because I chose badly-vectored words".
In 20$ a die, they could sell Gameboy style cartridges for different models.
Okay, now _this_ is the cyberpunk future I asked for.
That would be very cool, get an upgraded model every couple of months. Maybe PCIe form factor.
Yes, and even holding couple of cartridges for different scenarios e.g image generation, coding, tts/stt, etc
Make them shaped like floppy disks to confuse the younger generations.
Microsoft
dude that would be so incredibly cool
> Certainly interesting for very low latency applications which need < 10k tokens context.
I’m really curious if context will really matter if using methods like Recursive Language Models[0]. That method is suited to break down a huge amount of context into smaller subagents recursively, each working on a symbolic subset of the prompt.
The challenge with RLM seemed like it burned through a ton of tokens to trade for more accuracy. If tokens are cheap, RLM seems like it could be beneficial here to provide much more accuracy over large contexts despite what the underlying model can handle
0. https://arxiv.org/abs/2512.24601
Don’t forget that the 8B model requires 10 of said chips to run.
And it’s a 3bit quant. So 3GB ram requirement.
If they run 8B using native 16bit quant, it will use 60 H100 sized chips.
> Don’t forget that the 8B model requires 10 of said chips to run.
Are you sure about that? If true it would definitely make it look a lot less interesting.
Their 2.4 kW is for 10 chips it seems based on the next platform article.
I assume they need all 10 chips for their 8B q3 model. Otherwise, they would have said so or they would have put a more impressive model as the demo.
https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...
It doesn’t make any sense to think you need the whole server to run one model. It’s much more likely that each server runs 10 instances of the model
1. It doesn’t make sense in terms of architecture. It’s one chip. You can’t split one model over 10 identical hardwire chips
2. It doesn’t add up with their claims of better power efficiency. 2.4kW for one model would be really bad.
We are both wrong.
First, it is likely one chip for llama 8B q3 with 1k context size. This could fit into around 3GB of SRAM which is about the theoretical maximum for TSMC N6 reticle limit.
Second, their plan is to etch larger models across multiple connected chips. It’s physically impossible to run bigger models otherwise since 3GB SRAM is about the max you can have on an 850mm2 chip.
https://mlq.ai/news/taalas-secures-169m-funding-to-develop-a...Aren't they only using the SRAM for the KV cache? They mention that the hardwired weights have a very high density. They say about the ROM part:
> We have got this scheme for the mask ROM recall fabric – the hard-wired part – where we can store four bits away and do the multiply related to it – everything – with a single transistor. So the density is basically insane.
I'm not a hardware guy but they seem to be making a strong distinction between the techniques they're using for the weights vs KV cache
> In the current generation, our density is 8 billion parameters on the hard wired part of the chip., plus the SRAM to allow us to do KV caches, adaptations like fine tuning, and etc.
Thanks for having a brain.
Not sure who started that "split into 10 chips" claim, it's just dumb.
This is Llama 3B hardcoded (literally) on one chip. That's what the startup is about, they emphasize this multiple times.
It’s just dumb to think that one chip per model is their plan. They stated that their plan is to chain multiple chips together.
I was indeed wrong about 10 chips. I thought they would use llama 8B 16bit and a few thousand context size. It turns out, they used llama 8B 3bit with around 1k context size. That made me assume they must have chained multiple chips together since the max SRAM on TSMC n6 for reticle sized chip is only around 3GB.
Were we go towards really smart roboters. It is interesting what kind of diferent model chips they can produce.
There is nothing smart about current LLMs. They just regurgitate text compressed in their memory based on probability. None of the LLMs currently have actual understanding of what you ask them to do and what they respond with.
If LLMs just regurgitate compressed text, they'd fail on any novel problem not in their training data. Yet, they routinely solve them, which means whatever's happening between input and output is more than retrieval, and calling it "not understanding" requires you to define understanding in a way that conveniently excludes everything except biological brains.
I somewhat agree with you but I also realise that there are very few "novel" problems in the world. I think it's really just more complex problem spaces is all.
Same relative logic, just more of it/more steps or trials.
Yes there are some fascinating emergent properties at play, but when they fail it's blatantly obvious that there's no actual intelligence nor understanding. They are very cool and very useful tools, I use them on a daily basis now and the way I can just paste a vague screenshot with some vague text and they get it and give a useful response blows my mind every time. But it's very clear that it's all just smoke and mirrors, they're not intelligent and you can't trust them with anything.
When humans fail a task, it’s obvious there is no actual intelligence nor understanding.
Intelligence is not as cool as you think it is.
It can still be cool- but maybe it's just not as rare.
I assure you, intelligence is very cool.
you'd think with how often Opus builds two separate code paths without feature parity when you try to vibe code something complex, people wouldn't regard this whole thing so highly
> they'd fail on any novel problem not in their training data
Yes, and that's exactly what they do.
No, none of the problems you gave to the LLM while toying around with them are in any way novel.
None of my codebases are in their training data, yet they routinely contribute to them in meaningful ways. They write code that I'm happy with that improves the codebases I work in.
Do you not consider that novel problem solving?
Correct, you are not doing any novel problem solving.
They don't solve novel problems. But if you have such strong belief, please give us examples.
Depends how precisely you define novel - I don't think LLMs are yet capable of posing and solving interesting problems, but they have been used to address known problems, and in doing so have contributed novel work. Examples include Erdos Problem #728[0] (Terence Tao said it was solved "more or less autonomously" by an LLM), IMO problems (Deepmind, OpenAI and Huang 2025), GPT-5.2 Pro contributing a conjecture in particle physics[1], systems like AlphaEvolve leveraging LLMs + evolutionary algorithms to generate new, faster algorithms for certain problems[2].
[0] https://mathstodon.xyz/@tao/115855840223258103
[1] https://huggingface.co/blog/dlouapre/gpt-single-minus-gluons
[2] https://deepmind.google/blog/alphaevolve-a-gemini-powered-co...
We know that, but that does not make them unuseful. The opposite in fact, they are extremely useful in the hands of non-idiots.We just happen to have a oversupply of idiots at the moment, which AI is here to eradicate. /Sort of satire.
So you are saying they are like copy, LLMs will copy some training data back to you? Why do we spend so much money training and running them if they "just regurgitate text compressed in their memory based on probability"? billions of dollars to build a lossy grep.
I think you are confused about LLMs - they take in context, and that context makes them generate new things, for existing things we have cp. By your logic pianos can't be creative instruments because they just produce the same 88 notes.
I have a gut feeling, huge portion of deficiencies we note with AI is just reflection of the training data. For instance, wiki/reddit/etc internet is just a soup of human description of the world model, not the actual world model itself. There are gaps or holes in the knowledge because codified summary of world is what is remarkable to us humans, not a 100% faithful, comprehensive description of the world. What is obvious to us humans with lived real world experience often does not make it into the training data. A simple, demonstrable example is whether one should walk or drive to car wash.
Thats not how they work, pro-tip maybe don't comment until you have a good understanding?
Would you mind rectifying the wrong parts then?
Phrases like "actual understanding", "true intelligence" etc. are not conducive to productive discussion unless you take the trouble to define what you mean by them (which ~nobody ever does). They're highly ambiguous and it's never clear what specific claims they do or don't imply when used by any given person.
But I think this specific claim is clearly wrong, if taken at face value:
> They just regurgitate text compressed in their memory
They're clearly capable of producing novel utterances, so they can't just be doing that. (Unless we're dealing with a very loose definition of "regurgitate", in which case it's probably best to use a different word if we want to understand each other.)
The fact that the outputs are probabilities is not important. What is important is how that output is computed.
You could imagine that it is possible to learn certain algorithms/ heuristics that "intelligence" is comprised of. No matter what you output. Training for optimal compression of tasks /taking actions -> could lead to intelligence being the best solution.
This is far from a formal argument but so is the stubborn reiteration off "it's just probabilities" or "it's just compression". Because this "just" thing is getting more an more capable of solving tasks that are surely not in the training data exactly like this.
Huh? Their words are an accurate, if simplified, description of how they work.
The simplification is where it loses granularity. I could describe every human's life as they were born and then they died. That's 100% accurate, but there's just a little something lost by simplifying that much.
Just HI slop. Ask any decent model, it can explain what's wrong this this description.
> 880mm^2 die
That's a lot of surface, isn't it? As big an M1 Ultra (2x M1 Max at 432mm² on TSMC N5P), a bit bigger than an A100 (820mm² on TSMC N7) or H100 (814mm² on TSMC N5).
> The larger the die size, the lower the yield.
I wonder if that applies? What's the big deal if a few parameter have a few bit flips?
> I wonder if that applies? What's the big deal if a few parameter have a few bit flips?
We get into the sci-fi territory where a machine achieves sentience because it has all the right manufacturing defects.
Reminds me of this https://en.wikipedia.org/wiki/A_Logic_Named_Joe
Also see Adrian Thompson's Xilinx 6200 FPGA, programmed by a genetic algorithm that worked but exploited nuances unique to that specific physical chip, meaning the software couldn't be copied to another chip. https://news.ycombinator.com/item?id=43152877
I love that story.
2000s movie line territory:
> There have always been ghosts in the machine. Random segments of code, that have grouped together to form unexpected protocols.
An on-device reasoning model what that kind of speed and cost would completely change the way people use their computers. It would be closer to star trek than anything else we've ever had. You'd never have to type anything or use a mouse again.
Hardware decoders make sense for fixed codecs like MPEG, but I can't see it making sense for small models that improve every 6 months.
It's weird to me to train such huge models to then destroy them by using them a 3 bits quantization per presumably 16bits (bfloat16) weights. Why not just train smaller models then.
K-V caches are large, but hidden states aren't necessarily that large. And if you can run a model once ridiculously fast, then you can loop it repeatedly and still be fast. So I wonder about the 'modern RNNs' like RWKV here...
There’s a bit of a hidden cost here… the longevity of GPU hardware is going to be longer, it’s extended every time there’s an algorithmic improvement. Whereas any efficiency gains in software that are not compatible with this hardware will tend to accelerate their depreciation.
There is nothing new here. This has been demonstrated several times by previous researchers:
https://arxiv.org/abs/2511.06174
https://arxiv.org/abs/2401.03868
For a real world use case, you would need an FPGA with terabytes of RAM. Perhaps it'll be a Off chip HBM. But for s large models, even that won't be enough. Then you would need to figure out NV-link like interconnect for these FPGAs. And we are back to square one.
This is new. You are citing FPGA prototypes. Those papers do not demonstrate the same class of scaling or hardware integration that Taalas is advocating. For one, the FPGA solutions typically use fixed multipliers (or lookup tables), the ASIC solution has more freedom to optimize routing for 4 bit multiplication.
Do not overlook traditional irrational investor exuberance, we've got an abundance of that right now. With the right PR manouveurs these guys could be a tulip craze.
Maybe they can stack LLM parameters in 200 layers like 3D NAND flash and make the chip very small ...
This is insane if true - could be super useful for data extraction tasks. Sounds like we could be talking in the cents per millions of tokens range.
Yea its fast af but very quickly loses context/hallucinates from my own tests with large chunks of text
Doesn't the blog state that it's now 4bit (the first gen was 3bit + 6bit)?
Sounds perfect for use in consumer devices.
Low-latency inference is a huge waste of power; if you're going to the trouble of making an ASIC, it should be for dog-slow but very high throughput inference. Undervolt the devices as much as possible and use sub-threshold modes, multiple Vt and body biasing extensively to save further power and minimize leakage losses, but also keep working in fine-grained nodes to reduce areas and distances. The sensible goal is to expend the least possible energy per operation, even at increased latency.
Low latency inference is very useful in voice-to-voice applications. You say it is a waste of power but at least their claim is that it is 10x more efficient. We'll see but if it works out it will definitely find its applications.
This is not voice-to-voice though, end-to-end voice chat models (the Her UX) are completely different.
I haven't found any end-to-end voice chat models useful. I had much better results with separate STT-LLM-TTS. One big problem is the turn detection and having inference with 150-200ms latency would allow for a whole new level of quality. I would just use it with a prompt: "You think the user is finished talking?" and then push it to a larger model. The AI should reply within the ballpark of 600ms-1000ms. Faster is often irritating, slower will make the user to start talking again.
I think it's really useful for agent to agent communication, as long as context loading doesn't become a bottleneck. Right now there can be noticeable delays under the hood, but at these speeds we'll never have to worry about latency when chain calling hundreds or thousands of agents in a network (I'm presuming this is going to take off in the future). Correct me if I'm wrong though.
What's happening in the comment section? How come so many cannot understand that his is running Llama 3.1 8B? Why are people judging its accuracy? It's almost a 2 years old 8B param model, why are people expecting to see Opus level response!?
The focus here should be on the custom hardware they are producing and its performance, that is whats impressive. Imagine putting GLM-5 on this, that'd be insane.
This reminds me a lot of when I tried the Mercury coder model by Inceptionlabs, they are creating something called a dLLM which is like a diffusion based llm. The speed is still impressive when playing aroun with it sometimes. But this, this is something else, it's almost unbelievable. As soon as I hit the enter key, the response appears, it feels instant.
I am also curious about Taalas pricing.
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Do we have an idea of how much a unit / inference / api will cost?
Also, considering how fast people switch models to keep up with the pace. Is there really a potential market for hardware designed for one model only? What will they do when they want to upgrade to a better version? Throw the current hardware and buy another one? Shouldn't there be a more flexible way? Maybe only having to switch the chip on top like how people upgrade CPUs. I don't know, just thinking out loudly.
They don't give cost figures in their blog post but they do here:
https://www.nextplatform.com/wp-content/uploads/2026/02/taal...
Probably they don't know what the market will bear and want to do some exploratory pricing, hence the "contact us" API access form. That's fair enough. But they're claiming orders of magnitude cost reduction.
> Is there really a potential market for hardware designed for one model only?
I'm sure there is. Models are largely interchangeable especially as the low end. There are lots of use cases where you don't need super smart models but cheapness and fastness can matter a lot.
Think about a simple use case: a company has a list of one million customer names but no information about gender or age. They'd like to get a rough understanding of this. Mapping name -> guessed gender, rough guess of age is a simple problem for even dumb LLMs. I just tried it on ChatJimmy and it worked fine. For this kind of exploratory data problem you really benefit from mass parallelism, low cost and low latency.
> Shouldn't there be a more flexible way?
The whole point of their design is to sacrifice flexibility for speed, although they claim they support fine tunes via LoRAs. LLMs are already supremely flexible so it probably doesn't matter.
Yes, there are all kinds of fuzzy NLP tasks that this would be great for. Jobs where you can chunk the text into small units and add instructions and only need a short response. You could burn through huge data sets very quickly using these chips.
That is my concern too. A chip optimised for a model or specific model architecture will not be useful for long.
I just tried the demo and I think, this is huge! If they manage to build a chip in 2 or 3 years, that can run something like Opus 4.6 or even Sonnet, at that speed, the disruption in the world of software development will be more than we saw in the last 3-5 years. LLMs today are somewhat useful, but they are still too slow and expensive for a meaningful ralph loop. Being able to runs those loops (or if you want to call it "thinking") much faster, will enable a lot of stuff, that is not feasible today. Writing things like openclaw will not take weeks, but hours. Maybe even rewriting entire tools, kernels or OSes will be feasible because the LLM can run through almost endless tries.
Speed and cost wins over quality and this will also be true for LLMs.
I personally don't buy it, cerebras is way more advanced than this, comparing this tok/s to cerebras is disingenious.
Cerebras is a totally different product though. They can (theoretically) run any frontier model provided it gets compiled a certain way. Like a wafer scale TPU.
This is using hardwired weights with on-die SRAM used for K/V for example. It's WAY more power efficient and faster. The tradeoff being it's hardwired.
Still, most frontier models are "good enough" where an obscenely fast version would be a major seller.
If it's so easy to do custom silicon for any model (they say only 2 months), why didn't they demo one of the newer DeepSeek models instead? Using a 2-year model is so bad. I'm not buying it.
they explain it in the article: this is the first iteration, so they wanted to start with something simple, ie, this is a tech demo.
Ok then I look forward to seeing DeepSeek running instantly at the end of April.
Why so negative lol. The speed and very reduced power use of this thing are nothing to be sneezed at. I mean, hardware accelerated LLMs are a huge step forward. But yeah, this is a proof of concept, basically. I wouldn't be surprised if the size factor and the power use go down even more, and that we'll start seeing stuff like this in all kinds of hardware. It's an enabler.
You don't know. You just have marketing materials, not independent analysis. Maybe it actually takes 2 years to design and manufacture the hardware, so anything that comes out will be badly out of date. Wouldn't be the first time someone lied. A good demo backed by millions of dollars should not allow such doubts.
Did you not see the chatbot they posted online (https://chatjimmy.ai/)? That thing is near instantaneous, it's all the proof you need that this is real.
And if the hardware is real and functional, as you can independently verify by chatting with that thing, how much more effort would it be to etch more recent models?
The real question is of course: what about LARGER models? I'm assuming you can apply some of the existing LLM inference parallelization techniques and split the workload over multiple cards. Some of the 32B models are plenty powerful.
It's a proof of concept, and a convincing one.
They support Lora, it is something.
Holy cow their chatapp demo!!! I for first time thought i mistakenly pasted the answer. It was literally in a blink of an eye.!!
https://chatjimmy.ai/
I asked it to design a submarine for my cat and literally the instant my finger touched return the answer was there. And that is factoring in the round-trip time for the data too. Crazy.
The answer wasn't dumb like others are getting. It was pretty comprehensive and useful.
it's incredible how many people are commenting here without having read the article. they completely lost the point.
With this speed, you can keep looping and generating code until it passes all tests. If you have tests.
Generate lots of solutions and mix and match. This allows a new way to look at LLMs.
Not just looping, you could do a parallel graph search of the solution-space until you hit one that works.
Infinite Monkey Theory just reached its peak
You could also parse prompts into an AST, run inference, run evals, then optimise the prompts with something like a genetic algorithm.
Agreed, this is exciting, and has me thinking about completely different orchestrator patterns. You could begin to approach the solution space much more like a traditional optimization strategy such as CMA-ES. Rather than expect the first answer to be correct, you diverge wildly before converging.
And then it's slow again to finally find a correct answer...
It won't find the correct answer. Garbage in, garbage out.
How about if you run this loop (one year from now) on this kind of hardware but with something like Claude/Kimi K2. How about that? Because that's where it'll go.
This is what people already do with “ralph” loops using the top coding models. It’s slow relative to this, but still very fast compared to hand-coding.
This doesn't work. The model outputs the most probable tokens. Running it again and asking for less probable tokens just results in the same but with more errors.
Do you not have experience with agents solving problems? They already successfully do this. They try different things until they get a solution.
OK investors, time to pull out of OpenAI and move all your money to ChatJimmy.
A related argument I raised a few days back on HN:
What's the moat with with these giant data-centers that are being built with 100's of billions of dollars on nvidia chips?
If such chips can be built so easily, and offer this insane level of performance at 10x efficiency, then one thing is 100% sure: more such startups are coming... and with that, an entire new ecosystem.
RAM hoarding is, AFAICT, the moat.
lol... true that for now though
Yeah, just cause Cisco had a huge market lead on telecom in the late '90s, it doesn't mean they kept it.
(And people nowadays: "Who's Cisco?")
They did mostly keep it though.
Sure, but it's taken their stock price about 20 years to recover.
I think their hope is that they’ll have the “brand name” and expertise to have a good head start when real inference hardware comes out. It does seem very strange, though, to have all these massive infrastructure investment on what is ultimately going to be useless prototyping hardware.
Tools like openclaw start making the models a commodity.
I need some smarts to route my question to the correct model. I wont care which that is. Selling commodities is notorious for slow and steady growth.
You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training will still require general purpose GPU compute
Yeah but you need even bigger factories to fabricate those inference chips, so what is the point?
Next up: wire up a specialized chip to run the training loop of a specific architecture.
If I am not mistaken this chip was build specifically for the llama 8b model. Nvidia chips are general purpose.
Nvidia bought all the capacity so their competitors can't be manufactured at scale.
You mean Nvidia?
> It was literally in a blink of an eye.!!
It's not even close. It takes the eye 100mm .. 400ms to blink. This think takes under 30ms to process a small query - about 10 times faster.
I got 16.000 tokens per second ahaha
I dunno, it pretty quickly got stuck; the "attach file" didn't seem to work, and when I asked "can you see the attachment" it replied to my first message rather than my question.
It’s llama 3.1 8B. No vision, not smart. It’s just a technical demo.
why is everyone seemingly incapable of understanding this? waht is going on here? Its like ai doomers consistently have the foresight of a rat. yeah no shit it sucks its running llama 3 8b, but theyre completely incapable of extrapolation.
Hmm.. I had tried simple chat converation without file attachments.
Well it got all 10 incorrect when I asked for top 10 catchphrases from a character in Plato's books. It confused the baddie for Socrates.
Well yeah, they're running a small, outdated, older model. That's not really the point. This approach can be used for better, larger, newer models.
I get nothing, no replies to anything.
Maybe hn and reddit crowd have overloaded them lol
That… what…
I asked, “What are the newest restaurants in New York City?”
Jimmy replied with, “2022 and 2023 openings:”
0_0
Well, technically it's answer is correct when you consider it's knowledge cutoff date... it just gave you a generic always right answer :)
chatjimmy's trained on LLama 3.1
Is super fast but also super inaccurate, I would say not even gpt-3 levels.
That's because it's llama3 8b.
There are a lot of people here that are completely missing the point. What is it called where you look at a point of time and judge an idea without seemingly being able to imagine 5 seconds into the future.
“static evaluation”
It is incredibly fast, on that I agree, but even simple queries I tried got very inaccurate answers. Which makes sense, it's essentially a trade off of how much time you give it to "think", but if it's fast to the point where it has no accuracy, I'm not sure I see the appeal.
the hardwired model is Llama 3.1 8B, which is a lightweight model from two years ago. Unlike other models, it doesn't use "reasoning:" the time between question and answer is spent predicting the next tokens. It doesn't run faster because it uses less time to "think," It runs faster because its weights are hardwired into the chip rather than loaded from memory. A larger model running on a larger hardwired chip would run about as fast and get far more accurate results. That's what this proof of concept shows
I see, that's very cool, that's the context I was missing, thanks a lot for explaining.
I don't mean to be rude, but did you read the article before commenting?
I'm commenting on the link to their demo, not on the article.
If it's incredibly fast at a 2022 state of the art level of accuracy, then surely it's only a matter of time until it's incredibly fast at a 2026 level of accuracy.
yeah this is mindblowing speed. imagine this with opus 4.6 or gpt 5.2. probably coming soon
I'd be happy if they can run GLM 5 like that. It's amazing at coding.
Why do you assume this?
I can produce total jibberish even faster, doesn’t mean I produce Einstein level thought if I slow down
Better models already exist, this is just proving you can dramatically increase inference speeds / reduce inference costs.
It isn't about model capability - it's about inference hardware. Same smarts, faster.
Not what he said.
I think it might be pretty good for translation. Especially when fed with small chunks of the content at a time so it doesn't lose track on longer texts.
Fast, but stupid.
The question is not about how fast it is. The real question(s) are: (This also assumes that diffusion LLMs will get faster)The blog answers all those questions. It says they're working on fabbing a reasoning model this summer. It also says how long they think they need to fab new models, and that the chips support LoRAs and tweaking context window size.
I don't get these posts about ChatJimmy's intelligence. It's a heavily quantized Llama 3, using a custom quantization scheme because that was state of the art when they started. They claim they can update quickly (so I wonder why they didn't wait a few more months tbh and fab a newer model). Llama 3 wasn't very smart but so what, a lot of LLM use cases don't need smart, they need fast and cheap.
Also apparently they can run DeepSeek R1 also, and they have benchmarks for that. New models only require a couple of new masks so they're flexible.
The counting rs in strawberry problem was a example of people not understanding how the models work but I guess good to show the limitations of the current architectures.
But thing is, those architectures haven't improved a whole lot. Now when it answers that correctly it's either in training data or by virtue of "count letters" or code sandbox tools.
LLMs can't count. They need tool use to answer these questions accurately.
That particular one can't count without using external tools. Others can, and do.
[dead]
A lot of naysayers in the comments, but there are so many uses for non-frontier models. The proof of this is in the openrouter activity graph for llama 3.1: https://openrouter.ai/meta-llama/llama-3.1-8b-instruct/activ...
10b daily tokens growing at an average of 22% every week.
There are plenty of times I look to groq for narrow domain responses - these smaller models are fantastic for that and there's often no need for something heavier. Getting the latency of reponses down means you can use LLM-assisted processing in a standard webpage load, not just for async processes. I'm really impressed by this, especially if this is its first showing.
Maybe this is a naive question, but why wouldn't there be market for this even for frontier models? If Anthropic wanted to burn Opus 4.6 into a chip, wouldn't there theoretically be a price point where this would lower inference costs for them?
Because we don't know if this would scale well to high-quality frontier models. If you need to manufacture dedicated hardware for each new model, that adds a lot of expense and causes a lot of e-waste once the next model releases. In contrast, even this current iteration seems like it would be fantastic for low-grade LLM work.
For example, searching a database of tens of millions of text files. Very little "intelligence" is required, but cost and speed are very important. If you want to know something specific on Wikipedia but don't want to figure out which article to search for, you can just have an LLM read the entire English Wikipedia (7,140,211 articles) and compile a report. Doing that would be prohibitively expensive and glacially slow with standard LLM providers, but Taalas could probably do it in a few minutes or even seconds, and it would probably be pretty cheap.
These seem ideal for robotics applications, where there is a low-latency narrow use case path that these chips can serve, maybe locally.
Exactly. One easily relatable use-case is structured content extraction or/and conversion to markdown for web page data. I used to use groq for same (gpt-oss20b model), but even that used to feel slow when doing theis task at scale.
LLM's have opened-up natural language interface to machines. This chip makes it realtime. And that opens a lot of use-cases.
Many older models are still better at "creative" tasks because new models have been benchmarking for code and reasoning. Pre-training is what gives a model its creativity and layering SFT and RL on top tends to remove some of it in order to have instruction following.
I have such a deep need for something that's just a step above semantic search. These non-frontier models running blazingly fast can solve that.
So many problems simply don't require a full LLM, but more than traditional software. Training a novel model isn't really a compelling argument at most tech startups right now, so you need to find an LLM-native way to do things.
I've never gotten incorrect answers faster than this, wow!
Jokes aside, it's very promising. For sure a lucrative market down the line, but definitely not for a model of size 8B. I think lower level intellect param amount is around 80B (but what do I know). Best of luck!
Make it for Qwen 2.5 and I'd buy it.
You don't actually need "frontier models" for Real Work (c).
(Summarization, classification and the rest of the usual NLP suspects.)
I completely agree. So many things can benefit from having "smart classifiers".
Like, give me semantic search that can detect the difference between SSL and TLS without needing to put a full LLM in the loop.
As someone with a 3060, I can attest that there are really really good 7-9B models. I still use berkeley-nest/Starling-LM-7B-alpha and that model is a few years old.
If we are going for accuracy, the question should be asked multiple times on multiple models and see if there is agreement.
But I do think once you hit 80B, you can struggle to see the difference between SOTA.
That said, GPT4.5 was the GOAT. I can't imagine how expensive that one was to run.
Amazing! It couldn't answer my question at all, but it couldn't answer it incredibly quickly!
Snarky, but true. It is truly astounding, and feels categorically different. But it's also perfectly useless at the moment. A digital fidget spinner.
does no one understand what a tech demo is anymore? do you think this piece of technology is just going to be frozen in time at this capability for eternity?
do you have the foresight of a nematode?
Yeah, two p’s in the word pepperoni …
Edit: it seems like this is likely one chip and not 10. I assumed 8B 16bit quant with 4K or more context. This made me think that they must have chained multiple chips together since N6 850mm2 chip would only yield 3GB of SRAM max. Instead, they seem to have etched llama 8B q3 with 1k context instead which would indeed fit the chip size.
This requires 10 chips for an 8 billion q3 param model. 2.4kW.
10 reticle sized chips on TSMC N6. Basically 10x Nvidia H100 GPUs.
Model is etched onto the silicon chip. So can’t change anything about the model after the chip has been designed and manufactured.
Interesting design for niche applications.
What is a task that is extremely high value, only require a small model intelligence, require tremendous speed, is ok to run on a cloud due to power requirements, AND will be used for years without change since the model is etched into silicon?
I'm thinking the best end result would come from custom-built models. An 8 billion parameter generalized model will run really quickly while not being particularly good at anything. But the same parameter count dedicated to parsing emails, RAG summarization, or some other specialized task could be more than good enough while also running at crazy speeds.
> What is a task that is extremely high value, only require a small model intelligence, require tremendous speed, is ok to run on a cloud due to power requirements, AND will be used for years without change since the model is etched into silicon?
Video game NPCs?
Doesn’t pass the high value and require tremendous speed tests.
Video games are a huge market, and speed and cost of current models are definitely huge barriers to integrating LLMs in video games.
Speed = capacity = cost.
Alternatively, you could run far more RAG and thinking to integrate recent knowledge, I would imagine models designed for this putting less emphasis on world knowledge and more on agentic search.
Maybe; models with more embedded associations are also better at search. (Intuitively, this tracks; a model with no world knowledge has no awareness of synonyms or relations (a pure markov model), so the more knowledge a model has, the better it can search.) It’s not clear if it’s possible to build such a model, since there doesn’t seem to be a scaling cliff.
Where are those numbers from? It's not immediately clear to me that you can distribute one model across chips with this design.
> Model is etched onto the silicon chip. So can’t change anything about the model after the chip has been designed and manufactured.
Subtle detail here: the fastest turnaround that one could reasonably expect on that process is about six months. This might eventually be useful, but at the moment it seems like the model churn is huge and people insist you use this week's model for best results.
And what of that makes you assume that having a server with 10 HC1 cards is needed to run a single model version on that server?
So it lights money on fire extra fast, AI focused VCs are going to really love it then!!
Well they claim two month turnaround. Big If True. How does the six months break down in your estimation? Maybe they have found a way to reduce the turnaround time.
This depends on how much better the models will get from now in, if Claude Opus 4.6 was transformed into one of these chips and ran at a hypothetical 17k tokens/second, I'm sure that would be astounding, this depends on how much better claude Opus 5 would be compared to the current generation
Even an O3 quality model at that speed would be incredible for a great many tasks. Not everything needs to be claude code. Imagine Apple fine tuning a mid tier reasoning model on personal assistant/MacOs/IOS sorts of tasks and burning a chip onto the mac studio motherboard. Could you run claude code on it? Probably not, would it be 1000x better than Siri? absolutely.
Yeah, waiting for Apple to cut a die that can do excellent local AI.
I’m pretty sure they’d need a small data center to run a model the size of Opus.
100x of a less good model might be better than 1 of a better model for many many applications.
This isn't ready for phones yet, but think of something like phones where people buy new ones every 3 years and even having a mediocre on-device model at that speed would be incredible for something like siri.
Data tagging? 20k tok/s is at the point where I'd consider running an LLM on data from a column of a database, and these <=100 token problems provide the least chance of hallucination or stupidity.
A lot of NLP tasks could benefit from this
ceo
No one would never give such a weak model that much power over a company.
The NextPlatform article hints at their approach:
“We have got this scheme for the mask ROM recall fabric – the hard-wired part – where we can store four bits away and do the multiply related to it – everything – with a SINGLE TRANSISTOR. So the density is basically insane. And this is not nuclear physics – it is fully digital. It is just a clever trick that we don’t want to broadcast. But once you hardwire everything, you get this opportunity to stuff very differently than if you have to deal with changing things. The important thing is that we can put a weight and do the multiply associated with it all in one transistor. And you know the multipliers are kind of the big boy piece of the computer.“
One transistor doing 4-bit multiplication? A plausible way to get “4-bit weight plus multiply in one transistor” in a 6 nm FinFET mask-ROM fabric is to make the ROM cell a single device whose drive strength is the stored value. At tapeout you pick one of about 16 discrete strengths (for example by choosing fin count and possibly Vt), so that transistor itself encodes a 4-bit weight. Then you do the multiply in the charge/time domain by encoding the input activation as a discrete pulse width or pulse count and letting the cell source or sink a weight-proportional current onto a precharged bitline for that duration. The resulting bitline voltage change (or time-to-threshold) is proportional to current times time, so it behaves like weight times input and can be accumulated along a column before a simple comparator or time-to-digital readout. It’s “digital” in the sense that both weight and input are quantized, but it relies on device physics; the hard parts are keeping 16 levels separable across PVT, mismatch, and aging, plus managing bitline noise and coupling and ensuring the device stays in a predictable operating region.
VLSI design produces digital outputs, but in the quantum silicon domain, it’s all about the analog…
TIL your salary
(Kiddin’, my silly way to say thanks for a deeply technical look, helps me understand the kind of knowledge work that might be useful n years from now!)
There's an old idea of adaptive media. Imagine a video drama that's composed of a graph of clips, like an old "choose your own adventure" book ("Do you X? If yes, goto page 45"). With gaze tracking, one can "hmm, the viewer is more focused on character A than B... so we'll give clips and subplots with more A".
Now, when reading, the eye moves in little jumps - saccades. They last 10's of ms, the eye is blind during them, and with high-quality tracking, you know quite early just where that foveal peephole is going to land. So handwave a budget of a few ms for trajectory analysis, a few for 200 Hz rendering latency, and you still have 10-ish ms to play with. At 20k tok/s, that's 200 tok.
So perhaps one might JIT the next sentence, or the topic of the next paragraph, or the entire nature of the document, based on the user's attention. Imagine a universal document - you start reading, and you find the document is about, whatever you wanted it to be about?
Generative TikTok for words
Hmm... TikTok has apparently long had "text enhanced with background" genres, and TIL, text posts since 2023. So text is ok. But non-independent items? For generative storytelling, "here is a next paragraph for the story", swipe left/right might work? Want to avoid "I don't much like this new paragraph, but I'm afraid to lose it and be stuck with something worse". Swipe left/right and up for continue? Swipe down to revisit old choices? Maybe present new text bolded, appended to old text, for context. Or a "next page of a picture book" idiom. A text field for direct creative or editorial intervention - speech to text. Maybe a side channel input for "story and background should now be soporific". Generative bedtime stories, but incrementally collaboratively created... Thanks for the brainstorming prompt.
If I could have one of these cards in my own computer do you think it would be possible to replace claude code?
1. Assume It's running a better model, even a dedicated coding model. High scoring but obviously not opus 4.5 2. Instead of the standard send-receive paradigm we set up a pipeline of agents, each of whom parses the output of the previous.
At 17k/tps running locally, you could effectively spin up tasks like "you are an agent who adds semicolons to the end of the line in javascript", with some sort of dedicated software in the style of claude code you could load an array of 20 agents each with a role to play in improving outpus.
take user input and gather context from codebase -> rewrite what you think the human asked you in the form of an LLM-optimized instructional prompt -> examine the prompt for uncertainties and gaps in your understanding or ability to execute -> <assume more steps as relevant> -> execute the work
Could you effectively set up something that is configurable to the individual developer - a folder of system prompts that every request loops through?
Do you really need the best model if you can pass your responses through a medium tier model that engages in rapid self improvement 30 times in a row before your claude server has returned its first shot response?
Models can't improve themselves with their own (model) input, they need to be grounded in truth and reality.
But at one point the model is sufficiently large enough to accomplish any task a human could specify. For software development, I think we're pretty much at that point with the latest Anthropic/Google/OpenAI models. We have no idea where the direction of token pricing is going to go in the future, but the consensus seems to be that it will only get more expensive. If Taalas can offer the same functionality that we have with frontier models today at a 1/10 of the cost and 10x the speed then they're going to take over a large part of the market.
I think so. The last few months have shown us that it isn't necessarily the models themselves that provide good results, but the tooling / harness around it. Codex, Opus, GLM 5, Kimi 2.5, etc. all each have their quirks. Use a harness like opencode and give the model the right amount of context, they'll all perform well and you'll get a correct answer every time.
So in my opinion, in a scenario like this where the token output is near instant but you're running a lower tier model, good tooling can overcome the differences between a frontier cloud model.
It's 2.5kW so it likely won't sit in your computer (quite beyond what a desktop could provide in power alone to a single card, let alone cool). It's 8.5cm^2 which is a beast of a single die.
Basically logistically it's going to need to be in a data centre.
It's ideal for small context high throughput. Perhaps parsing huge text piles like if you had the entire Epstein files as text.
I think Claude code benefits from larger context to keep your entire project in view and deep reasoning.
What this would certainly replace is when Claude dispatched to Haiku for manual NLP tasks.
> It's 2.5kW so it likely won't sit in your computer (quite beyond what a desktop could provide in power alone to a single card, let alone cool). It's 8.5cm^2 which is a beast of a single die.
I wonder how you cool a 3x3cm die that outputs 2.5 kW of heat. In the article they mention that the traditional setup requires water cooling, but surely this does as well, right?
Can't imagine what else could manage that nearly 2.8W/mm2.
It does make you wonder if they copy is misleading about something so simple how much else could be puffery?
Maybe they mean that a standard liquid cooling system will work?
I tried the chatbot. jarring to see a large response come back instantly at over 15k tok/sec
I'll take one with a frontier model please, for my local coding and home ai needs..
Absolute insanity to see a coherent text block that takes at least 2 minutes to read generated in a fraction of a second. Crazy stuff...
Accelerating the end of the usable text-based internet one chip at a time.
Not at all if you consider the internet pre-LLM. That is the standard expectation when you load a website.
The slow word-by-word typing was what we started to get used to with LLMs.
If these techniques get widespread, we may grow accustomed to the "old" speed again where content loads ~instantly.
Imagine a content forest like Wikipedia instantly generated like a Minecraft word...
Yes, but the quality of the output leaves to be desired. I just asked about some sports history and got a mix of correct information and totally made up nonsense. Not unexpected for an 8k model, but raises the question of what the use case is for such small models.
8b models are great at converting unstructured data to a structured format. Say, you want to transcribe all your customer calls and get a list of issues they discussed most often. Currently with the larger models it takes me hours.
A chatbot which tells you various fun facts is not the only use case for LLMs. They're language models first and foremost, so they're good at language processing tasks (where they don't "hallucinate" as much).
Their ability to memorize various facts (with some "hallucinations") is an interesting side effect which is now abused to make them into "AI agents" and what not but they're just general-purpose language processing machines at their core.
Would be nice to point this at (pre-LLM) Wikipedia and fill out Wikidata!
You have a misunderstanding of what LLMs are good at.
Poster wants it to play Jeopardy, not process text.
Not sure if you're correct, as the market is betting trillions of dollars on these LLMs, hoping that they'll be close to what the OP had expected to happen in this case.
The market didn't throw trillions of dollars to develop Llama 3 8B.
What GP is expected to happen has happened around late 2024 ~ early 2025 when LLM frontends got web search feature. It's old tech now.
The GP’s point was about LLMs generally, no matter the interface. I agree that this particular model is (relatively speaking) ancient in AI the world, but go back 3 or 4 years and this (pretty complex “reasoning” at almost instant speed) would have seemed taken out of a science-fiction book.
I don't think he does. Larger models are definitely better at not hallucinating. Enough that they are good at answering questions on popular topics.
Smaller models, not so much.
Care to enlighten me?
Don't ask a small LLM about precise minutiae factual information.
Alternatively, ask yourself how plausible it sounds that all the facts in the world could be compressed into 8k parameters while remaining intact and fine-grained. If your answer is that it sounds pretty impossible... well it is.
Did you see the part in my original post where it said "Not unexpected for an 8k model"?
Oh I saw it, you still have a fundamentally flawed comprehension of LLMs.
The size of the model does not factor as tiny models can use Internet to fetch factual information.
But you think they are accurate repositories of knowledge, even though it's physically impossible unless lossless infinite compression algorithms exist (they don't, can't and won't).
I think you're overestimating your ability to assess what others think or comprehend.
Reminds me of that solution to Fermi's paradox, that we don't detect signals from extraterrestrial civilizations because they run on a different clock speed.
Iain M Banks’ The Algebraist does a great job of covering that territory. If an organism had a lifespan of millions of years, they might perceive time and communication differently to say a house fly or us.
:eyeroll:
Yeah, feeding that speed into a reasoning loop or a coding harness is going to revolutionize AI.
So cool, what's underappreciated imo: 17k tokens/sec doesn't just change deployment economics. It changes what evaluation means, static MMLU-style tests were designed around human-paced interaction. At this throughput you can run tens of thousands of adversarial agent interactions in the time a standard benchmark takes. Speed doesn't make static evals better it makes them even more obviously inadequate.
The speed of the chatbot's response is startling when you're used to the simulated fast typing of ChatGPT and others. But the Llama 3.1 8B model Taalas uses predictably results in incorrect answers, hallucinations, poor reliability as a chatbot.
What type of latency-sensitive applications are appropriate for a small-model, high-throughput solution like this? I presume this type of specialization is necessary for robotics, drones, or industrial automation. What else?
Coding, for some future definition of "small-model" that expands to include today's frontier models. What I commented a few days ago on codex-spark release:
"""
We're going to see a further bifurcation in inference use-cases in the next 12 months. I'm expecting this distinction to become prominent:
(A) Massively parallel (optimize for token/$)
(B) Serial low latency (optimize for token/s).
Users will switch between A and B depending on need.
Examples of (A):
- "Use subagents to search this 1M line codebase for DRY violations subject to $spec."
An example of (B):
- "Diagnose this one specific bug."
- "Apply these text edits".
(B) is used in funnels to unblock (A).
"""
You could build realtime API routing and orchestration systems that rely on high quality language understanding but need near-instant responses. Examples:
1. Intent based API gateways: convert natural language queries into structured API calls in real time (eg., "cancel my last order and refund it to the original payment method" -> authentication, order lookup, cancellation, refund API chain).
2. Of course, realtime voice chat.. kinda like you see in movies.
3. Security and fraud triage systems: parse logs without hardcoded regexes and issue alerts and full user reports in real time and decide which automated workflows to trigger.
4. Highly interactive what-if scenarios powered by natural language queries.
This effectively gives you database level speeds on top of natural language understanding.
Routing in agent pipelines is another use. "Does user prompt A make sense with document type A?" If yes, continue, if no, escalate. That sort of thing
For this type of repetitive application I think it's common to "fine-tune" a model trained on your business problem to reach higher quality/reliability metrics. That might not be possible with this chip.
They say LoRA finetunes work.
I'm wondering how much the output quality of a small model could be boosted by taking multiple goes at it. Generate 20 answers and feed them back through with a "rank these responses" prompt. Or doing something like MCTS.
Isn't this what thinking models do internally? Chain of thoughts?
No. Chain of thought it just the model generating a single answer for longer inside <think></think> tags which are not shown in the final response. The strategy of generating different answers in parallel is something different (which can be used in conjunction with chain of thought) and is the thing used by models like Gemini 3 Deep Think and GPT-5.2 Pro.
Hmm.. got it. Thanks..
Maybe summarization? I’d still worry about accuracy but smaller models do quite well.
Language translation, chunk by chunk.
> Jimmy: What is the largest planet in our solar system?
> me: the moon
> Jimmy: The answer to "What is the capital of France?" I was looking for was the city of Paris, but that's not the correct response to the original question of the capital of France. The question that got cut off was actually "What is the capital of France?", and the response "There are plenty of times I look to groq for narrow domain responses" wasn't the answer I was looking for.
It is certainly fast, but I think there might be some caching issues somewhere.
The speed is ridiunkulous. No doubt.
The quantization looks pretty severe, which could make the comparison chart misleading. But I tried a trick question suggested by Claude and got nearly identical results in regular ollama and with the chatbot. And quantization to 3 or 4 bits still would not get you that HOLY CRAP WTF speed on other hardware!
This is a very impressive proof of concept. If they can deliver that medium-sized model they're talking about... if they can mass produce these... I notice you can't order one, so far.
I doubt many of us will be able to order one for a long while. There is a significant number of existing datacentre and enterprise use-cases that will pay a premium for this.
Additionally LLMs have been tested, found valuable in benchmarks, but not used for a large number of domains due to speed and cost limitations. These spaces will eat up these chips very quickly.
I wonder if this makes the frontier labs abandon the SAAS per-token pricing concept for their newest models, and we'll be seeing non-open-but-on-chip-only models instead, sold by the chip and not by the token.
It could give a boost to the industry of electron microscopy analysis as the frontier model creators could be interested in extracting the weights of their competitors.
The high speed of model evolution has interesting consequences on how often batches and masks are cycled. Probably we'll see some pressure on chip manufacturers to create masks more quickly, which can lead to faster hardware cycles. Probably with some compromises, i.e. all of the util stuff around the chip would be static, only the weights part would change. They might in fact pre-make masks that only have the weights missing, for even faster iteration speed.
This would be killer for exploring simultaneous thinking paths and council-style decision taking. Even with Qwen3-Coder-Next 80B if you could achieve a 10x speed, I'd buy one of those today. Can't wait to see if this is still possible with larger models than 8B.
It uses 10 chips for 8B model. It’d need 80 chips for an 80b model.
Each chip is the size of an H100.
So 80 H100 to run at this speed. Can’t change the model after you manufacture the chips since it’s etched into silicon.
As many others in this conversation have asked, can we have some sources on the idea that the model is spread across chips? You keep making the claim, but no one (myself included) else has any idea where that information comes from or if it is correct.
I was indeed wrong about 10 chips. I thought they would use llama 8B 16bit and a few thousand context size. It turns out, they used llama 8B 3bit with only 1k context size. That made me assume they must have chained multiple chips together since the max SRAM on TSMC n6 for reticle sized chip is only around 3GB.
I'm sure there is plenty of optimization paths left for them if they're a startup. And imho smaller models will keep getting better. And a great business model for people having to buy your chips for each new LLM release :)
One more thing. It seems like this is a Q3 quant. So only 3GB RAM requirement.
10 H100 chips for 3GB model.
I think it’s a niche of a niche at this point.
I’m not sure what optimization they can do since a transistor is a transistor.
Do we know that it needs 10 chips to run the model? Or are the servers for the API and chatbot just specced with 10 boards to distribute user load?
If you etch the bits into silicon, you then have to accommodate the bits by physical area, which is the transistor density for whatever modern process they use. This will give you a lower bound for the size of the wafers.
Wow I'm impressed. I didn't actually think we'd see it encoded on chips. Or well I knew some layer of it could be, some sort of instruction set and chip design but this is pretty staggering. It opens the door to a lot of things. Basically it totally destroys the boundaries of where software will go but I also think we'll continue to see some generic chips show up that hit this performance soon enough. But the specialised chips with encoded models. This could be what ends up in specific places like cars, planes, robots, etc where latency matters. Maybe I'm out of the loop, I'm sure others and doing it including Google.
Reminds me of when bitcoin started running on ASICs. This will always lag behind the state of the art, but incredibly fast, (presumably) power efficient LLMs will be great to see. I sincerely hope they opt for a path of selling products rather than cloud services in the long run, though.
17k TPS is slow compared to other probabilistic models. It was possible to hit ~10-20 million TPS decades ago with n-gram and PDFA models, without custom silicon. A more informative KPI would be Pass@k on a downstream reasoning task - for many such benchmarks, increasing token throughput by several orders of magnitude does not even move the needle on sample efficiency.
Bunch of negative sentiment in here, but I think this is pretty huge. There are quite a lot of applications where latency is a bigger requirement than the complexity of needing the latest model out there. Anywhere you'd wanna turn something qualitative into something quantitative but not make it painfully obvious to a user that you're running an LLM to do this transformation.
As an example, we've been experimenting with letting users search free form text, and using LLMs to turn that into a structured search fitting our setup. The latency on the response from any existing model simply kills this, its too high to be used for something where users are at most used to the delay of a network request + very little.
There are plenty of other usecases like this where.
This is what’s gonna be in the brain of the robot that ends the world.
The sheer speed of how fast this thing can “think” is insanity.
try here, I hate llms but this is crazy fast. https://chatjimmy.ai/
The full answer pops in milliseconds, it's impressive and feels like a completely different technology just by foregoing the need to stream the output.
Because most models today generate slowish, they give the impression of someone typing on the other end. This is just <enter> -> wall of text. Wild
We need that for this chinese 3B model that think 45s for hello world but also solves math.
Nanbeige. Yeah this seems ideal for models that scale test time compute
Do we know anything about the method?
Pretty cool, what they need is to build a tool that can take any model to chip in short a time as possible. How quick can they give me DeepSeek, Kimi, Qwen or GLM on a chip? I'll take 5k tk/sec for those!
also imagine it will cost 300$/unit, we all will host our own set of models locally, dream dream
I was wondering if/when this would happen. My friends and I would discuss this at the pub all the time, "LLM2RTL" or take it a step further and do the the whole process "LLM2GDS".
I couldn't find much info here, but I'm guessing they've built tooling to automatically convert model weights to RTL and the reason it's such an old model is that it takes a long time tape a chip out (especially the first one). Would be interesting to know how much is automated and how much is handcrafted.
I think the "next big thing" with AI hardware will be when they switch from "digital" implementations of LLMs to "analogue". We already know that we can lose some bits of precision and still have a "workable" model. If/when folks figure the fine-tuning out, I'm guessing it'll be another order of magnitude improvement.
Token velocity is great, but the industry is hyper-fixated on speed while completely ignoring the blast radius. If we push to 17k tokens/sec for autonomous agents, we are just accelerating how fast an agent can hit an infinite loop and drain an API budget. Before we make AI ubiquitous, we need deterministic, network-level circuit breakers. Speed without governance is just a faster way to burn capital.
I wanted to try the demo so I found the link
> Write me 10 sentences about your favorite Subway sandwich
Click button
Instant! It was so fast I started laughing. This kind of speed will really, really change things
If it's not reprogrammable, it's just expensive glass.
If you etch the bits into silicon, you then have to accommodate the bits by physical area, which is the transistor density for whatever modern process they use. This will give you a lower bound for the size of the wafers.
This can give huge wafers for a very set model which is old by the time it is finalized.
Etching generic functions used in ML and common fused kernels would seem much more viable as they could be used as building blocks.
Models don’t get old as fast as they used to. A lot of the improvements seem to go into making the models more efficient, or the infrastructure around the models. If newer models mainly compete on efficiency it means you can run older models for longer on more efficient hardware while staying competitive.
If power costs are significantly lower, they can pay for themselves by the time they are outdated. It also means you can run more instances of a model in one datacenter, and that seems to be a big challenge these days: simply building an enough data centres and getting power to them. (See the ridiculous plans for building data centres in space)
A huge part of the cost with making chips is the masks. The transistor masks are expensive. Metal masks less so.
I figure they will eventually freeze the transistor layer and use metal masks to reconfigure the chips when the new models come out. That should further lower costs.
I don’t really know if this makes sanse. Depends on whether we get new breakthroughs in LLM architecture or not. It’s a gamble essentially. But honestly, so is buying nvidia blackwell chips for inference. I could see them getting uneconomical very quickly if any of the alternative inference optimised hardware pans out
From my own experience, models are at the tipping point for being useful at prototypes in software, and those are very large frontier models not feasible to get down on wafers unless someone does something smart.
I really don't like the hallucination rate for most models but it is improving, so that is still far in the future.
What I could see though, is if the whole unit they made would be power efficient enough to run on a robotics platform for human computer interaction.
It makes sense they would try to make repurposing their tech as much as they could since making changes is frought with a long time frame and risk.
But if we look long term and pretend that they get it to work, they just need to stay afloat until better smaller models can be made with their technology, so it becomes a waiting game for investors and a risk assessment.
> From my own experience, models are at the tipping point for being useful at prototypes in software
You must not have much experience using the new frontier models then. A lot of large tech companies are replacing their SDLC with agentic workflows. The tooling and frameworks are still ramping up, but the models have no problem producing production ready software given proper specifications.
“ Models don’t get old as fast as they used to”
^^^ I think the opposite is true
Anthropic and OpenAI are releasing new versions every 60-90 days it seems now, and you could argue they’re going to start releasing even faster
Are they becoming better at the same rate as before though?
In my unscientific experience, yes, but being better at a certain rate is hard to really quantify, unless you just pull some random benchmark numbers.
Per release, I’d say no.
Per period of time, I’d say yes.
No.
yes, pretty much
Reading the in depth article also linked in this thread, they say that only 2 layers need to change most of the time. They claim from new model to PCB in 2 months. Let's see, but sounds promising.
You don’t need it to be reprogrammable if it can use tools and RAG.
I'm curious how much of "hardcoding" is in the chip? Can it have parts that don't need changing much and "offload" the rest into some sort of high-speed/bandwidth interconnect?
Will we reach a state where we have chips on which models can be "flashed" like CPU firmware?
Or eventually will we reach a state where none of these tricks will be needed because like run-of-the mill Intel/AMD commodity CPUs, we will have full-power AI chips which will be part of an bigger/integrated mother-chip? Then what will happen to companies that do LLMs-as-a-service? Will they be forced to join and adapt becoming hybrid model+hardware shops?
I'm not knowledgeable enough about hardware but throwing these random ideas out in hopes of thought-provoking responses to learn from.
These chips are large by fab standards and even with state of the art processes we likely won't see any kind of integration on consumer tech any time soon, but I imagine they will absolutely see instant demand if they can deliver on what they laid out in the post.
>Founded 2.5 years ago, Taalas developed a platform for transforming any AI model into custom silicon. From the moment a previously unseen model is received, it can be realized in hardware in only two months.
So this is very cool. Though I'm not sure how the economics work out? 2 months is a long time in the model space. Although for many tasks, the models are now "good enough", especially when you put them in a "keep trying until it works" loop and run them at high inference speed.
Seems like a chip would only be good for a few months though, they'd have to be upgrading them on a regular basis.
Unless model growth plateaus, or we exceed "good enough" for the relevant tasks, or both. The latter part seems quite likely, at least for certain types of work.
On that note I've shifted my focus from "best model" to "fastest/cheapest model that can do the job". For example testing Gemini Flash against Gemini Pro for simple tasks, they both complete the task fine, but Flash does it 3x cheaper and 3x faster. (Also had good results with Grok Fast in that category of bite-sized "realtime" workflows.)
My concept was to do this with two pieces:
1. Generic, mask layers and board to handle what's common across models. Especially memory and interface.
2. Specific layers for the model implementation.
Masks are the most expensive part of ASIC design. So, keeping the custom part small with the rest pre-proven in silicon, even shared across companies, would drop the costs significantly. This is already done in hardware industry in many ways but not model acceleration.
Then, do 8B, 30-40B, 70B, and 405B models in hardware. Make sure they're RLHF-tuned well since changes will be impossible or limited. Prompts will drive most useful functionality. Keep cranking out chips. There's maybe a chance to keep the weights changeable on-chip but it should still be useful if only inputs can change.
The other concept is to use analog, neural networks with the analog layers on older, cheaper nodes. We only have to customize that per model. The rest is pre-built digital with standard interfaces on a modern node. Given the chips would be distributed, one might get away with 28nm for the shared part and develop it eith shuttle runs.
I always thought eventually someone would come along and make a hardware accelerator for LLMs, but I thought it would be like google TPUs where you can load up whatever model you want. Baking the model into hardware sounds like the monkey paw curled, but it might be interesting selling an old.. MPU..? because it wasn't smart enough for your latest project
This is an interesting piece of hardware though when they go multi-chip for larger models the speed will no doubt suffer.
They'll also be severely limited on context length as it needs to sit in SRAM. Looks like the current one tops out at 6144 tokens which I presume is a whole chips worth. You'd also have to dedicate a chip to a whole user as there's likely only enough SRAM for one user's worth of context. I wonder how much time it takes them to swap users in/out? I wouldn't be surprised if this chip is severely underutilized (can't use it all when running decode as you have to run token by token with one users and then idle time as you swap users in/out).
Maybe a more realistic deployment would have chips for linear layers and chips for attention? You could batch users through the shared weight chips and then provision more or less attention chips as you want which would be per user (or shared amongst a small group 2-4 users).
Wow, this is great.
To the authors: do not self-deprecate your work. It is true this is not a frontier model (anymore) but the tech you've built is truly impressive. Very few hardware startups have a v1 as good as this one!
Also, for many tasks I can think of, you don't really need the best of the best of the best, cheap and instant inference is a major selling point in itself.
I am thinking if this can be a low-level substrate for composing dumb LLMs into smart swarm, theoretically:
1. A whole with disparate parts (smart and dumb components) are almost always more cost-effective to reach a given target of performance 2. With that, a whole with disparate parts, are almost always more performant with the same cost
A few inspiration:
1. Human body is intelligent composed of so diverse parts 2. Swarm intelligence of insects and small animals are certainly beyond current understanding
The cost and speed of this thing is on point to make such a whole composed diverse parts possible.
There are so many use cases for small and super fast models that are already in size capacity -
* Many top quality tts and stt models
* Image recognition, object tracking
* speculative decoding, attached to a much bigger model (big/small architecture?)
* agentic loop trying 20 different approaches / algorithms, and then picking the best one
* edited to add! Put 50 such small models to create a SOTA super fast model
The model routing discussion is fascinating. We're seeing similar patterns in how startups approach global talent - the best solution depends heavily on context. For engineering teams distributed across timezones (we work a lot with LATAM developers), the real bottleneck isn't just the model or the tool, it's understanding when to apply what. Same with inference: a 16k tok/s chip is incredible for real-time voice agents, but most startup use cases don't need that latency. The interesting question is whether we'll see more specialized hardware for niche applications, or if general-purpose solutions will keep winning through sheer volume economics.
I think the thing that makes 8b sized models interesting is the ability to train unique custom domain knowledge intelligence and this is the opposite of that. Like if you could deploy any 8b sized model on it and be this fast that would be super interesting, but being stuck with llama3 8b isn't that interesting.
The "small model with unique custom domain knowledge" approach has a very low capability ceiling.
Model intelligence is, in many ways, a function of model size. A small model tuned for a given domain is still crippled by being small.
Some things don't benefit from general intelligence much. Sometimes a dumb narrow specialist really is all you need for your tasks. But building that small specialized model isn't easy or cheap.
Engineering isn't free, models tend to grow obsolete as the price/capability frontier advances, and AI specialists are less of a commodity than AI inference is. I'm inclined to bet against approaches like this on a principle.
> Engineering isn't free, models tend to grow obsolete as the price/capability frontier advances, and AI specialists are less of a commodity than AI inference is. I'm inclined to bet against approaches like this on a principle.
This does not sound like it will simplify the training and data side, unless their or subsequent models can somehow be efficiently utilized for that. However, this development may lead to (open source) hardware and distributed system compilation, EDA tooling, bus system design, etc getting more deserved attention and funding. In turn, new hardware may lead to more training and data competition instead of the current NVIDIA model training monopoly market. So I think you're correct for ~5 years.
A fine tuned 1.7B model probably is still too crippled to do anything useful. But around 8b the capabilities really start to change. I’m also extremely unemployed right now so I can provide the engineering.
Asking it what its knowledge cut-off is interesting, it doesn't seem to be consistent even within a single response. Sometimes it responds to say it cuts off 2020 too.
The instantaneous response is impressive though. I'm sure there will be applications for this, I just lack the imagination to know what they'll be.Sounds like they aren’t providing a system prompt which tells it what its knowledge cutoff is. It doesn’t know unless it’s told. Given the limited context, it makes sense they wouldn’t use up valuable space for that.
The system prompt is configurable in the sidebar and minimal. It doesn't give a knowledge cutoff. This is a tech demo of the fact it works at all, it's not meant to be a good chatbot.
So they create a new chip for every model they want to support, is that right? Looking at that from 2026, when new large models are coming out every week, that seems troubling, but that's also a surface take. As many people here know better than I that a lot of the new models the big guys release are just incremental changes with little optimization going into how they're used, maybe there's plenty of room for a model-as-hardware model.
Which brings me to my second thing. We mostly pitch the AI wars as OpenAI vs Meta vs Claude vs Google vs etc. But another take is the war between open, locally run models and SaaS models, which really is about the war for general computing. Maybe a business model like this is a great tool to help keep general computing in the fight.
We’re reaching a saturation threshold where older models are good enough for many tasks, certainly at 100x faster inference speeds. Llama3.1 8B might be a little too old to be directly useful for e.g. coding but it certainly gets the gears turning about what you could do with one Opus orchestrator and a few of these blazing fast minions to spit out boilerplate…
One of these things, however old, coupled with robust tool calling is a chip that could remain useful for decades. Baking in incremental updates of world knowledge isn't all that useful. It's kinda horrifying if you think about it, this chip among other things contains knowledge of Donald Trump encoded in silicon. I think this is a way cooler legacy for Melania than the movie haha.
Minor note to anyone from taalas:
The background on your site genuinely made me wonder what was wrong with my monitor.
The company slogan is great: "The Model is The Computer"
It's an homage to Jensen: "The display is the computer"
https://www.wired.com/2002/07/nvidia/
This is incredible. With this speed I can use LLMs in a lot of pre-filtering etc. tasks. As a trivial example, I have a personal OpenClaw-like bot that I use to do a bunch of things. Some of the things just require it to do trivial tool-calling and tell me what's up. Things like skill or tool pre-filtering become a lot more feasible if they're always done.
Anyway, I imagine these are incredibly expensive, but if they ever sell them with Linux drivers and slotting into a standard PCIe it would be absolutely sick. At 3 kW that seems unlikely, but for that kind of speed I bet I could find space in my cabinet and just rip it. I just can't justify $300k, you know.
The demo was so fast it highlighted a UX component of LLMs I hadn’t considered before: there’s such a thing as too fast, at least in the chatbot context. The demo answered with a page of text so fast I had to scroll up every time to see where it started. It completely broke the illusion of conversation where I can usually interrupt if we’re headed in the wrong direction. At least in some contexts, it may become useful to artificially slow down the delivery of output or somehow tune it to the reader’s speed based on how quickly they reply. TTS probably does this naturally, but for text based interactions, still a thing to think about.
That is what self-driving car should eventually use, whenever they (or the authorities) deem their model good enough. Burn it on a dedicated chip. It would be cheaper (energy) to run, and faster to make decisions.
It's more expensive in COGS and self driving doesn't need to run at 900 Hz.
wonder if at some point you could swap the model as if you were replacing a cpu in your pc or inserting a game cartridge
The number six seven
> It seems like "six seven" is likely being used to represent the number 17. Is that correct? If so, I'd be happy to discuss the significance or meaning of the number 17 with you.
This is genuinely an incredible proof-of-concept; the business implications of this demo to the AI labs and all the companies that derive a ton of profit from inference is difficult to understate, really.
I think this is how I'm going to get my dream of Opus 3.7 running locally, quickly and cheaply on my mid-tier MacBook in 2030. Amazing. Anthropic et al will be able to make marginal revenue from licensing the weights of their frontier-minus-minus models to these folks.
I do like the idea of an aftermarket of ancient LLM chips that still have tons of useful life on text processing tasks etc. They don't talk about their architecture much, I wonder how well power can scale down. 200W for such a small model is not something I see happening in a laptop any time soon. Pretty hilarious implications for moat-building of the big providers too.
Yea I mean this is the first publishable draft of a startup cooking on this.
I'm confident there are at least 1-2 OOMs of improvement to come here in terms of the (intelligence : wattage) ratio.
I really thought we were going to need to see a couple of dramatic OOM-improvement changes to the model composition / software layer, in order to get models of Opus 3.7's capability running on our laptops.
This release tells me that eventual breakthrough won't even be strictly necessary, imo.
The way I imagine it in 2-4 years we're going to be hit with a triple glut of better architecture, massive oversupply of hardware and potentially one or two hardware efforts like this really taking off. It's pretty crazy we're already 4 years in and outside of very niche / low availability solutions, it's still either GPU or bust
That's interesting! How do you see "oversupply of hardware" playing out?
Is it because we stop doing ~2024-style, large-scale training (marginal returns aren't worth it)? Or because supply way outpaces the training+inference demand?
AFAIU if the trend lines /S-curves keep chugging along as they are, we won't hit hardware oversupply for a long, long time without some sort of AI training winter.
I think frontier models can do more with fewer tokens (and do the wrong thing far less often) than a "really fast" small model.
There are use cases for fast/ultrafast inferrence models - classifying text, scoring things, extracting information - but for coding and other knowledge tasks - you're not going to get to your solution faster at 16,000 tokens/s if the solution never comes (or is the wrong one).
I think this is quite interesting for local AI applications. As this technology basically scales with parameter size, if there could be some ASIC for a QWen 0.5B or Google 0.3B model thrown onto a laptop motherboard it'd be very interesting.
Obviously not for any hard applications, but for significantly better autocorrect, local next word predictions, file indexing (tagging I suppose).
The efficiency of such a small model should theoretically be great!
Are the model weights burned into the silicon / part of the architecture? Or can you update the model weights on these chips? If they cannot be updated, these chips will be outdated the moment they are made given the breakneck speed at which new and improved models are introduced.
Applied for access. I hope to test a parallel fast-inference problem solver with a hybrid MCTS approach.
The implications for RLM is really interesting. RLM is expensive because of token economics. But when tokens are so cheap and fast to generate, context size of the model matters a lot less
Also interesting implications for optimization-driven frameworks like DSPy. If you have an eval loop and useful reward function, you can iterate to the best possible response every time and ignore the cost of each attempt
17k token/sec is $0.18/chip/hr for the size of H100 chip if they want to compete with the market rate[1]. But 17k token/sec could lead to some new usecases.
[1]: https://artificialanalysis.ai/models/llama-3-1-instruct-8b/p...
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Insane gains, makes me excited for the future. Imagine Opus-like responses in <1 second.
I suspect power efficiency will be nearly entirely offset by increased usage but it’s more bang for watt.
I am extremely impressed by their inference speed!
Embedding the model at chip fab time ought to be useful for robotics, driving, vision, and audio applications, at least. The training sets are good for years.
So they use 3 bit values. Is that current thinking? LLMs started at 32-bit floats, and have gradually shrunk. 8-bit floats seem to work. Is 3 bits pushing it?
If you're making your own chip, you might as well explore analog computation.
This is really cool! I am trying to find a way to accelerate LLM inference for PII detection purposes, where speed is really necessary as we want to process millions of log lines per minute, I am wondering how fast we could get e.g. llama 3.1 to run on a conventional NVIDIA card? 10k tokens per second would be fantastic but even at 1k this would be very useful.
PII redaction is a really good use-case.
Also, "10k tokens per second would be fantastic" might not be sufficient (even remotely) if you want to "process millions of log lines per minute".
Assuming a single log line at just 100 tokens, you need (100 * 2 million / 60) ~ 3.3 million tokens per second processing speed :)
Yeah I mean we have a mechanism that can bypass AI models for log lines where we are pretty sure no PII is in there (kind of like smart caching using fuzzy template matching to identify things that we have seen before many times, as logs tend to contain the same stuff over and over with tiny variations e.g. different timestamps), so we only need to pass the lines where we cannot be sure there's nothing to the AI for inspection. And we can of course parallelize. Currently we use a homebrew CFR model with lots of tweaks and it's quite good but an LLM would of course be much better still and capture a lof of cases that would evade the simpler model.
Oh okay... that's fine. Most log lines are indeed similar looking.
For that you only need high throughput which is much easier to achieve compared to high latency, thanks to batching -- assuming the log lines or chunks can be processed independently. You can check TensorRT-LLM benchmarks (https://nvidia.github.io/TensorRT-LLM/developer-guide/perf-o...), or try running vllm on a card you have access to.
Their "chat jimmy" demo sure is fast, but it's not useful at all.
Test prompt: ```
Please classify the sentiment of this post as "positive", "neutral" or "negative":
Given the price, I expected very little from this case, and I was 100% right.
``` Jimmy: Neutral.
I tried various other examples that I had successfully "solved" with very early LLMs and the results were similarly bad.
Maybe its the tism but I also read that sentence as neutral. You expected very little and you got very little. Why would that be positive or negative? Maybe it should be positive because you got what you were expecting? But I would call getting what you expect something neutral, if you expected little and got a lot then that would be positive. If you expected a lot and got little then its negative. But if you expected little and got little the most clear outcome is that its a neutral statement. Am I missing something?
> Taalas’ silicon Llama achieves 17K tokens/sec per user, nearly 10X faster than the current state of the art, while costing 20X less to build, and consuming 10X less power.
Am I reading this right: 10x faster and 10x less power, ie. 100x more power efficient?
Jarring to see these other comments so blindly positive.
Show me something at a model size 80GB+ or this feels like "positive results in mice"
There are a lot of problems solved by tiny models. The huge ones are fun for large programming tasks, exploration, analysis, etc. but there's a massive amount of processing <10GB happening every day. Including on portable devices.
This is great even if it can't ever run Opus. Many people will be extremely happy about something like Phi accessible at lightning speed.
Positive results in mice also known as being a promising proof of concept. At this point, anything which deflates the enormous bubble around GPUs, memory, etc, is a welcome remedy. A decent amount of efficient, "good enough" AI will change the market very considerably, adding a segment for people who don't need frontier models. I'd be surprised if they didn't end up releasing something a lot bigger than they have.
Parameter density is doubling every 3-4 months
What does that mean for 8b models 24mo from now?
This is crazy! These chips could make high-reasoning models run so fast that they could generate lots of solution variants and automatically choose the best. Or you could have a smart chip in your home lab and run local models - fast, without needing a lot of expensive hardware or electricity
If they made a low power/mobile version, this could be really huge for embedded electronics. Mass produced, highly efficient "good enough" but still sort of dumb ais could put intelligence in house hold devices like toasters, light switches, and toilets. Truly we could be entering into the golden age of curses.
Oh god, this is the new version of every device having Bluetooth and an app and being called “smart”.
I just wanted some toast, but here I am installing an app, dismissing 10 popups, and maybe now arguing with a chat bot about how I don’t in fact want to turn on notifications.
I totally buy the thesis on specialization here, I think it makes total sense.
Asides from the obvious concern that this is a tiny 8B model, I'm also a bit skeptical of the power draw. 2.4 kW feels a little bit high, but someone else should try doing the napkin math compared to the total throughput to power ratio on the H200 and other chips.
This makes me think about how large would an FPGA-based system to be able to do this? Obviously there is no single-chip FPGA that can do this kind of job, but I wonder how many we would need.
Also, what if Cerebras decided to make a wafer-sized FPGA array and turned large language models into lots and lots of logical gates?
one step closer to being able to purchase a box of llms on aliexpress, though 1.7ktok/s would be quite enough
This is pretty wild! Only Llama3.1-8B, but this is only their first release so you can assume they're working on larger versions.
So what's the use case for an extremely fast small model? Structuring vast amounts of unstructured data, maybe? Put it in a little service droid so it doesn't need the cloud?
I imagine how advantageous it would be to have something like llama.cpp encoded on a chip instead, allowing us to run more than a single model. It would be slower than Jimmy, for sure, but depending on the speed, it could be an acceptable trade-off.
I am super happy to see people working on hardware for local llm. Yet, isnt it premature ? Space is still evolving. Today, i refuse to buy a gpu because i do not know what will be the best model tomorrow. Waiting to get a on the shelf device to run an opus like model
Strange that they apparently raised $169M (really?) and the website looks like this. Don't get me wrong: Plain HTML would do if "perfect", or you would expect something heavily designed. But script-kiddie vibe coded seems off.
The idea is good though and could work.
Strange that they raised money at all with an idea like this.
It's a bad idea that can't work well. Not while the field is advancing the way it is.
Manufacturing silicon is a long pipeline - and in the world of AI, one year of capability gap isn't something you can afford. You build a SOTA model into your chips, and by the time you get those chips, it's outperformed at its tasks by open weights models half their size.
Now, if AI advances somehow ground to a screeching halt, with model upgrades coming out every 4 years, not every 4 months? Maybe it'll be viable. As is, it's a waste of silicon.
Poverty of imagination here, plenty uses of this and its a prototype at this stage.
What uses, exactly?
The prototype is: silicon with a Llama 3.1 8B etched into it. Today's 4B models already outperform it.
Token rate in five digits is a major technical flex, but, does anyone really need to run a very dumb model at this speed?
The only things that come to mind that could reap a benefit are: asymmetric exotics like VLA action policies and voice stages for V2V models. Both of which are "small fast low latency model backed by a large smart model", and both depend on model to model comms, which this doesn't demonstrate.
In a way, it's an I/O accelerator rather than an inference engine. At best.
With LLMs this fast, you could imagine using them as any old function in programs.
You could always have. Assuming you have an API or a local model.
Which was always the killer assumption, and this changes little.
Even if this first generation is not useful, the learning and architecture decisions in this generation will be. You really can't think of any value to having a chip which can run LLMs at high speed and locally for 1/10 of the energy budget and (presumably) significantly lower cost than a GPU?
If you look at any development in computing, ASICs are the next step. It seems almost inevitable. Yes, it will always trail behind state of the art. But value will come quickly in a few generations.
maybe they're betting on improvement in models to plateau, and that having a fairly stablized capable model that is orders of magnitude faster than running on GPU's can be valuable in the future?
the non-frontier model use case is underrated here. for things like autocomplete, classification, routing, or first-pass triage -- you don't need GPT-4 class reasoning. you need something that's fast and cheap enough to call on every keystroke or every incoming message. at 17k tok/sec and 20x lower cost, a whole class of latency-sensitive edge applications becomes viable that aren't today.
But as models are changing rapidly and new architectures coming up, how do they scale and also we do t yet know the current transformer architecture will scale more than it already is. Soo many ope questions but VCs seems to be pouring money.
It would be pretty incredible if they could host an embedding model on this same hardware, I would pay for that immediately. It would change the type of things you could build by enabling on the fly embeddings with negligible latency.
Imagine a mass-produced AI chips with all human knowledge packed in chinesium epoxy blobs running from CR2032 batteries in toys for children. Given the progress in density and power consumption, it's not that far away.
The token throughput improvements are impressive. This has direct implications for usage-based billing in AI products — faster inference means lower cost per request, which changes the economics of credits-based pricing models significantly.
I wonder if this is the first step towards AI as an appliance rather than a subscription?
So I'm guessing this is some kind of weights as ROM type of thing? At least that's how I interpret the product page, or maybe even a sort of ROM type thing that you can only access by doing matrix multiplies.
You shouldn't need any ROM. It's likely the architecture is just fixed hardware with weights loaded in via scan flip-flows. If it was me making it, I'd just design a systolic array. Just multipliers feeding into multipliers, without even going through RAM.
That seems promising for applications that require raw speed. Wonder how much they can scale it up - 8B model quantized is very usable but still quite small compared to even bottom end cloud models.
Can it scale to an 800 billion param model? 8B parameter models are too far behind the frontier to be useful to me for SWE work.
Or is that the catch? Either way I am sure there will be some niche uses for it.
Spam. :P
so 90% of the AI market?
Would it make sense for the big players to buy them? Seems to be a huge avenue here to kill inference costs which always made me dubious on LLMs in general.
Performance like that may open the door to the strategy of brutefocing solutions to problems for which you have a verifier (problems such as decompilation).
For fun I'm imagining a future where you would be able to buy an ASIC with like an hard-wired 1B LLM model in it for cents and it could be used everywhere.
We need one of these things running an OSS vision model. Having super-fast agentic computer access would be so worthwhile!
Fast but the output is shit due to the contrained model they used. Doubt we'll ever get something like this for the large Param decent models.
Yep, this is the most exciting demo for me yet. Holy cow this is unbelievably fast.
The most impressive demo since gpt 3, honestly.
Since we already have open source models that are plenty good, like the new kimi k2.5, all I need is the ability to run it at moderate speed.
Honestly I am not bullish on capabilities that models do not yet have, seems we have seen it all and the only advancement have been context size.
And honestly I would claim this is the market sentiment aswell, anthropic showed opus 4.6 first and the big release was actually sonnet, the model people would use routinely. Nobody gave a shit about Gemini 3.1 pro, 3.0 flash was very successful...
Given all the recent developments in the last 12 months, no new use cases have opened for me. Given this insane speed, even on a limited model/context size, we would approach IA very differently.
Amazing speed. Imagine if its standardised like the GPU card equivalent in the future.
New models come out, time to upgrade your AI card, etc.
I was all praise for Cerberus, and now this ! $30 M for PCIe card in hand, really makes it approachable for many startups.
Wow. I’m finding it hard to even conceive of what it’d be like to have one of the frontier models on hardware at this speed.
Is amazingly fast but since the model is quantized and pretty limited, I don't know what it is useful for.
This is like microcontrollers, but for AI? Awesome! I want one for my electric guitar; and please add an AI TTS module...
No, it's ASICs, but for AI.
This is impressive. If you can scale it to larger models, and somehow make the ROM writeable, wow, you win the game.
Inference is crazy fast! I can see lot of potential for this kind of chip for IOT devices and Robotics.
Future is these as small, swappable bits of SD-card sized hardware that you stick into your devices.
Gemini Flash 2.5 lite does 400 tokens/sec. Is there benefit to going faster than a person can read?
There is also the use case of delegating tasks programmatically to an LLM, for example, transforming unstructured data to structured data. This task often can’t be done reliably without either 1. lots of manual work, or 2. intelligence, especially when the structure of the individual data pieces are unknown. Problems like these can be much more efficiently solved by LLMs, and if you imagine these programs are processing very large datasets, then sub-millisecond inference is crucial.
Aren't such tasks inherently parrallelizable?
Agents also "read", so yes there is. Think about spinning up 10, 20, 100 sub agents for a small task and they all return near instant. That's the usecase, not the chatbot.
Agents already bypass human inference time, if it can input-output instantly it can also loop it generating near instantly long cached tasks
Yes. You can allow multiple people to use a single chip. A slower solution will be able to service far fewer users.
Right, but it is also possible it's cheaper to use 42 Google TPUs for a second than one of these.
Talks about ubiquitous AI but can't make a blog post readable for humans :/
That animated background is terrible.
Incredibly distracting. No way to turn it off (at least within what's provided without using something like devtools.)
I don't know the use of this yet but I'm certain there will be one.
I know it is not easy to see the benefits of small models easily but this is what I am building for (1). I created a product for Google Gemini 3 Hackathon and I used Gemini 3 Flash (2). I tested locally using Ministral 3B and it was promising. Definitely will need work. But 8B/14B may give awesome results.
I am building a data extraction software on top of emails, attachments, cloud/local files. I use a reverse template generation with only variable translation done by LLMs (3). Small models are awesome for this (4).
I just applied for API access. If privacy policies are a fit, I would love to enable this for MVP launch.
1. https://github.com/brainless/dwata
2. https://youtu.be/Uhs6SK4rocU
3. https://github.com/brainless/dwata/tree/feature/reverse-temp...
4. https://github.com/brainless/dwata/tree/feature/reverse-temp...
I still believe this is the right - and inevitable - path for AI, especially as I use more premium AI tooling and evaluate its utility (I’m still a societal doomer on it, but even I gotta admit its coding abilities are incredible to behold, albeit lacking in quality).
Everyone in Capital wants the perpetual rent-extraction model of API calls and subscription fees, which makes sense given how well it worked in the SaaS boom. However, as Taalas points out, new innovations often scale in consumption closer to the point of service rather than monopolized centers, and I expect AI to be no different. When it’s being used sparsely for odd prompts or agentically to produce larger outputs, having local (or near-local) inferencing is the inevitable end goal: if a model like Qwen or Llama can output something similar to Opus or Codex running on an affordable accelerator at home or in the office server, then why bother with the subscription fees or API bills? That compounds when technical folks (hi!) point out that any process done agentically can instead just be output as software for infinite repetition in lieu of subscriptions and maintained indefinitely by existing technical talent and the same accelerator you bought with CapEx, rather than a fleet of pricey AI seats with OpEx.
The big push seems to be building processes dependent upon recurring revenue streams, but I’m gradually seeing more and more folks work the slop machines for the output they want and then put it away or cancel their sub. I think Taalas - conceptually, anyway - is on to something.
> Though society seems poised to build a dystopian future defined by data centers and adjacent power plants, history hints at a different direction. Past technological revolutions often started with grotesque prototypes, only to be eclipsed by breakthroughs yielding more practical outcomes.
…for a privileged minority, yes, and to the detriment of billions of people whose names the history books conveniently forget. AI, like past technological revolutions, is a force multiplier for both productivity and exploitation.
What would it take to put Opus on a chip? Can it be done? What’s the minimum size?
Maybe not today. Opus is quite large. This demo works with a very small 8B model. But, maybe one day. Hopefully soon. Opus on a chip would be very awesome, even if it can never be upgraded.
Someone mentioned that maybe we'd see a future where these things come in something like Nintendo cartridges. Want a newer model? Pop in the right catridge.
I have a hard time reading beyond factual lies like:
> On the cost front, deploying modern models demands massive engineering and capital: room-sized supercomputers consuming hundreds of kilowatts…
This is just wrong. The largest models are probably 1-2 trillion parameters. Say 2 trillion and let’s pretend it’s only quantized to 8bit (even though it could easily be half that.) So we need 2TB of VRAM. Not even using the latest hardware, lets say H100 chips with 80GB of vram each, with 8 of them in say an 8U. (Although you can certainly fit these in 6U still air cooled or even 4U water cooled.) Three of these server would almost do, but let’s call it four to include plenty of room for context. The largest physical size would be 32U - most of a single rack. Which is hardly the size of a room, even in Manhattan. Total power maybe 40kW. And you could easily drop these numbers to a half or quarter of that with reasonable modifications or upgrades.
If you want to sell your hardware, start by being honest about the problem you’re addressing.
It was so fast that I didn't realise it had sent its response. Damn.
Hurrah, its dumb answer to the now classic "the car wash is 100m away, should I drive or walk?" appeared very quickly.
It's an 8B parameter model from a good while ago, what were your expectations?
"Many believe AI is the real deal. In narrow domains, it already surpasses human performance. Used well, it is an unprecedented amplifier of human ingenuity and productivity."
Sounds like people drinking the Kool-Aid now.
I don't reject that AI has use cases. But I do reject that it is promoted as "unprecedented amplifier" of human xyz anything. These folks would even claim how AI improves human creativity. Well, has this been the case?
For me, this is entirely true.
I'm progressing with my side projects like I've never before.
Same, I would have given up on them long ago, I no longer code at all now. Why would I when the latest models can do it better, faster and without the human limitations of tiredness, emotional impacts etc.
> These folks would even claim how AI improves human creativity. Well, has this been the case?
Yes. Example: If you've never programmed in language X, but want to build something in it, you can focus on getting from 0 to 1 instead of being bogged down in the idiosyncrasies of said language.
> These folks would even claim how AI improves human creativity. Well, has this been the case?
For many of us, the answer is an emphatic yes.
ASIC inference is clearly the future just as ASIC bitcoin mining was
Not sure, but is this just ASICs for a particular model release?
I'm loving summarization of articles using their chatbot! Wow!
Deep Differentiable Logic Gate Networks?
There's a scifi story here when millions of these chips, with Qwen8-AGI-Thinking baked into them, are obsoleted by the release of Qwen9-ASI, which promptly destroys humanity and then itself by accident. A few thousand years later, some of the Qwen8 chips in landfill somehow power back up again and rebuild civilization on Earth.
Paging qntm...
Does anyone have an idea how much such a component costs?
Is this hardware for sale ? The site doesn't say.
This is, so far, utterly charming. I made a simple prompt of "make an adventure game in the style of cia.bas from pc-sig". It ended up being wildly different than that, but 30 minutes later and I'm still busy trying to play this "game" it fabricated out of thin air.
One interesting thing is it keeps randomly emitting "ประก" (meaning "Announcement") and chartInstance. This is recalling the early days of GPT-2 when the light bulb went on that "hey, there's something groundbreaking here".
What happened to Beff Jezos AI Chip?
Jesus, it just generated a story in 0.039s.
Whoever doesn’t buy/replicate this in the next year is dead. Imagine OpenAI trying to sell you a platform that takes 15 minutes, when someone else can do it in 0.001s.
> generated a story
Was the story any good? The model used isn't comparable to something OpenAI would offer you today. What good is a fast gibberish generator?
Question:
> Generate a fiction book about an old sorcerer who, like all elderly magicians, has to go back to studying magic for two years at a magic school in order to be granted the right to keep his wand and to keep doing magic at his home.
Answer, generated in 0,063s @ 15 742 tok/s:
> The Relearning of Eryndor Thorne
> In the quaint village of Brindlemark, where ancient traditions and mystical energies coexisted, Eryndor Thorne had lived a life of quiet reverence. For nigh on five decades, he had wielded the arcane arts with precision and mastery, earning him the respect of his peers and the awe of the villagers. His home, a humble abode atop a hill overlooking the rolling hills and verdant valleys, was a sanctuary of sorts, a place where the very fabric of magic was woven into the air. His library was a treasure trove of ancient tomes, texts, and scrolls, filled with the wisdom of the ages. Yet, despite his mastery over the arcane, Eryndor felt the stirrings of a nagging concern. The once-renewable license to practice magic, a privilege granted by the Council of Mages, was on the cusp of expiration.
> As all elderly mages knew, every two decades, the Council would demand that those who had spent a lifetime mastering the arcane arts return to the prestigious Wysteria Academy, the hallowed halls of magic learning, to relearn the foundational principles of magic. This was a ritual as old as the craft itself, a test of the elderly mage's understanding, to prove they still possessed the fundamental knowledge and skills required to practice magic with precision and safety.
> Eryndor, like many of his fellow mages, had grown complacent in his mastery. He had mastered the intricacies of elemental magic, bending the winds, waters, and flames to his will. His wand, a family heirloom passed down through generations, had become an extension of his arm, channeling the power of magic with ease. Yet, despite his proficiency, the Council's decree was clear: Eryndor must return to Wysteria Academy for two years of reeducation.
> The journey was not an easy one. Packing his worn leather satchel with the few books and scrolls he deemed essential, Eryndor bid farewell to his loved ones in Brindlemark, leaving behind the familiarity of his home to embark on a journey that would take him to a place both wondrous and treacherous. At Wysteria Academy, he joined a cohort of students from all corners of the realm, each as skilled and eager as he once was. The relearning process was a grueling one, as if facing a mountain he had once conquered. New techniques were presented, new theories to unravel, and new expectations to meet. Eryndor, once a master, now sat at the feet of his teachers, absorbing the new knowledge like a parched traveler at an oasis.
> Among his fellow students, there was the young, fire-kissed mage, Elara, who wielded magic with an intensity that bordered on reckless abandon. Her fiery nature and quick wit often put her at odds with the strict, ancient traditions, earning her a certain notoriety among the academy's elder mages. Then there was the enigmatic, shadow-drawn Kael, whose mastery of the arcane was matched only by his mystery. Kael's affinity for the dark arts raised more than a few eyebrows among the faculty, but Eryndor, having once walked the fine line between light and shadow, saw something of himself in the young mage.
> As the years passed, Eryndor grew to appreciate the challenges and opportunities that came with his return to the academy. He found himself grappling with the nuances of magic anew, rekindling memories of his early days as a novice. The relearning process was as much about rediscovering himself as it was about mastering the arcane. His studies were a journey of self-discovery, one that tested the mettle of his will and the depths of his understanding.
> Upon completion of his studies, Eryndor stood before the Council once more, his wand in hand, his heart afire with the thrill of rediscovery. The Council's examination was not merely a test of his knowledge but a test of his character. Eryndor, like many of his peers, had grown complacent, but the rigors of relearning had rekindled a spark within him, a flame that would guide him through the trials ahead.
> With his renewed license granted, Eryndor returned to Brindlemark, his home and his heart rejuvenated. His library, once a testament to his mastery, was now a canvas for the new knowledge he had acquired. His wand, now an extension of his rekindled passion for the arcane, channelled magic with a precision and flair that was unmatched. Eryndor Thorne, the elderly mage, had not only relearned magic but had rediscovered himself, a testament to the transformative power of education and the enduring spirit of the arcane.
I don't know why, but my ultra wide monitor absolutely hates that site. The whole screen is flickering trying to deal with the annoying background. Thank the gods for reader mode.
he numbers being presented are deliberately misleading. On this model, Groq delivers around 1,300 tokens per second, whereas Cerebras achieves roughly 2,500 tokens per second.
With the next generation of Cerebras chips expected to be 5–7× faster, peak throughput could reach the ~17,500 tokens-per-second range. For smaller models like this, that level of performance is entirely realistic. So no, a general-purpose accelerator will likely continue to outperform a fixed-function ASIC with a specific model etched into it.
Moreover, we’re only looking at results from a two-year-old, relatively small model. We still don’t know how this architecture will scale with a large MoE model, especially given constraints like limited on-chip KV cache and more complex attention mechanisms.
The real test isn’t performance on a small benchmark model, it’s how the system handles large-scale, production-grade workloads under architectural constraints.
this is absolutely mindblowing speed. imagine this with opus or 5.2
It's crazily fast. But 8B model is pretty much useless.
Anyway VCs will dump money onto them, and we'll see if the approach can scale to bigger models soon.
Impressive stuff.
write six seven as a number
> The number "six" is actually a noun, not a number. However, I assume you're asking to write the number 7 as a numeral, which is: 7
The demo is dogshit: https://chatjimmy.ai/
I asked it some basic questions and it fudged it like it was chatgpt 1.0
Imagine this thing for autocomplete.
I'm not sure how good llama 3.1 8b is for that, but it should work, right?
Autocomplete models don't have to be very big, but they gotta be fast.
Holy shit this is fast. It generated a legible, original, two-paragraph story on given topics in 0.025s.
wow that is fast!
holy fuck this is really gud. imagine this with sota models. we are cooked af.damn
Holy Hallucinations, Batman...
User You: What is the airspeed of an unladen swallow/
Jimmy:
A reference to the classic line from Monty Python and the Holy Grail!
Thus far it is correct but here it comes:
In the movie, a group of medieval scholars, the Knights who say "Ni!" (also known as the Bridgekeepers), ask a king's servant, a blacksmith, if an unladen swallow makes good time. The blacksmith, played by John Cleese, responds with the now-famous line:
"And what's the airspeed velocity of an unladen swallow?"
To which the blacksmith responds:
"What do you mean? An unladen swallow?"
The blacksmith then explains that an unladen swallow is when a swallow (a type of bird) is flying in the air without carrying any burden, implying that it is moving at a speed that is not affected by carrying any weight or load.
Nope, that is not at all how it went. No blacksmiths, not related to the Knights who say 'Ni!' - who have nothing scholarly about them, they wanted a shrubbery and demanded another one when their wish was fulfilled. It was the Bridgekeeper - not related to the aforementioned knights - who asked 3 questions and would throw anyone who could not answer all three into the chasm. When he asked Arthur - played by Cleese - the question about the airspeed of the swallow he asked the Bridgekeeper whether he meant an African or European swallow. The Bridgekeeper did not have an answer and was thrown into the chasm, problem solved.
However, in reality, swallows are not typically "unladen" in flight. They are small birds that fly at relatively fast speeds, usually around 10-15 km/h (6-9 mph), but they are not usually carrying any burdens!
Needless LLM-blabber.
The "airspeed velocity of an unladen swallow" has become a meme and a cultural reference point, often used humorously or ironically to refer to situations where someone is trying to make an absurd or non-sensical argument or ask an absurd question.
Somewhat correct but not necessary in this context.
The correct answer to the question would have been Do you mean an African or European swallow? followed by a short reference to the movie.
Of course this demo is not about the accuracy of the model - 'an old Llama' as mentioned elsewhere in this thread - but it does show that speed isn't everything. For generating LLM-slop this hardware implementation probably offers an unbeatable price/performance ratio but it remains to be seen if it can be combined with larger and less hallucination-prone models.
> Holy Hallucinations, Batman...
Congratulations! You figured out that this is a demo of a very small 8B model from 2022.
Is it already available to buy, or is this a “pay now, get it later” kind of new ASIC miner? Sorry for being skeptical, but AI is the new "crypto coin", and the crypto bros are still around.
There doesn't seem to be any form of buying the HC1 hardware at the moment. There is a free chatbot demo and then a form to request access to API. They seem to intend HC1 to be for demonstration and HC2 for "real" use, but they don't seem to be taking payment for either at the moment.
[dead]
Scale this then close the loop and have fabs spit out new chips with latest weights every week that get placed in a server using a robot, how long before AGI?
The article doesn't say anything about the price (it will be expensive), but it doesn't look like something that the average developer would purchase.
An LLM's effective lifespan is a few months (ie the amount of time it is considered top-tier), it wouldn't make sense for a user to purchase something that would be superseded in a couple of months.
An LLM hosting service however, where it would operate 24/7, would be able to make up for the investment.
I tried the trick question I saw here before, about the make 1000 with 9 8s and additions only
I know it's not a resonating model, but I keep pushing it and eventually it gave me this as part of it's output
888 + 88 + 88 + 8 + 8 = 1060, too high... 8888 + 8 = 10000, too high... 888 + 8 + 8 +ประก 8 = 1000,ประก
I googled the strange symbol, it seems to mean Set in thai?
I don't think it's very valuable to talk about the model here, the model is just an old Llama. It's the hardware that matters.