Spaces:
Running
Please add newer models (deepseek v3)
Going by the benchmarks it's the best open-source model as of rn. It's a huge model but the active parameters are just 37B (which is cute compared to 70B of llama).
Benchmarks:
Live bench:
It's a "thinking" model so you'll also need a little addition to the UI to hide the thinking part.
Should I expect it to be available in hugging chat any time soon or is it too costly?
Either ways, I appreciate what you guys are doing and wish you the best.
Deepseek V3 is obviously a very big and spicy release. It will Certainly be very useful. BUT! It's - like - SOOO large, I don't believe that the HF team will host this 600B+ model.
They hosted Llama3.1 405B for a while, after which they took it down, since it simply consumed too many resources while also being a free service.
Deepseek V3 is a MoE model though, so it generates a LOT quicker than Llama 405B.
If we are lucky, maybe we get a Q4 quant or maybe even Q8 quant of the model on HuggingChat, but that would still eat over 300GB of VRAM, which is crazy expensive.