The path to ubiquitous AI (17k tokens/sec)

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

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...

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.

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/

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?

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.

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!

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?

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.

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

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.

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.

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..

> 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 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?

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 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!

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.

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.

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.

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.

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.

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!

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.

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

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.

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.

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

try here, I hate llms but this is crazy fast. https://chatjimmy.ai/

>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.)

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.

I am extremely impressed by their inference speed!

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.

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

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).

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 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/

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 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.

Minor note to anyone from taalas:

The background on your site genuinely made me wonder what was wrong with my monitor.

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.

What happened to Beff Jezos AI Chip?

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.

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.

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).

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

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.

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.

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?