35 points | by hochmartinez 2 days ago ago
13 comments
This is a demo of Taalas inference ASIC hardware. Prior discussion @ https://news.ycombinator.com/item?id=47086181
- https://news.ycombinator.com/item?id=47086181
- https://taalas.com/the-path-to-ubiquitous-ai/
- https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...
What model and hardware powers this?
Is this a Google T5 based model?
3bit hard-wired Llama 3.1 8B ( https://taalas.com/the-path-to-ubiquitous-ai/ )
3bit is a bit ridiculous. From that page I am unclear if the current model is 3 or 4bit. If it’s 4bit… well, NVIDIA showed that a well organized model can perform almost as well as 8bit.
I love seeing optimised SLM inference. Is there a current use-case for this? Edge CNNs make sense to me but not edge SLMs (yet).
If this is possible, why not all online AI engines work like this?
This is an specific model (Llama 3.1 8B) baked in hardware form. You can only use this model but get "low" power consumption and crazy speed.
If you want to run a different model you need new hardware for that new model.
It is really a crazy speed. 15k tokens/second.
I have tried it again. This is the future of chat UI, imho.
Generated in 0,074s • 15 754 tok/s
imagine a model like opus 4.6 at that speed, that would be insane
Impressive, but this particular underlying LLM is objectively weak. I'd like to see it done with a larger and newer better model.
This is a demo of Taalas inference ASIC hardware. Prior discussion @ https://news.ycombinator.com/item?id=47086181
- https://news.ycombinator.com/item?id=47086181
- https://taalas.com/the-path-to-ubiquitous-ai/
- https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...
What model and hardware powers this?
Is this a Google T5 based model?
3bit hard-wired Llama 3.1 8B ( https://taalas.com/the-path-to-ubiquitous-ai/ )
3bit is a bit ridiculous. From that page I am unclear if the current model is 3 or 4bit. If it’s 4bit… well, NVIDIA showed that a well organized model can perform almost as well as 8bit.
I love seeing optimised SLM inference. Is there a current use-case for this? Edge CNNs make sense to me but not edge SLMs (yet).
If this is possible, why not all online AI engines work like this?
This is an specific model (Llama 3.1 8B) baked in hardware form. You can only use this model but get "low" power consumption and crazy speed.
If you want to run a different model you need new hardware for that new model.
It is really a crazy speed. 15k tokens/second.
I have tried it again. This is the future of chat UI, imho.
Generated in 0,074s • 15 754 tok/s
imagine a model like opus 4.6 at that speed, that would be insane
Impressive, but this particular underlying LLM is objectively weak. I'd like to see it done with a larger and newer better model.