Hacker News Reader

Comment by ilaksh | Hacker News Reader

They just released a much larger model as a torrent. Probably the first open source mixture-of-experts model. Waiting for them to drop the code and explain it.

View on HN

Which one? Everything on their website is still about the 7B model. 7B is the only model available on their huggingface page too: https://huggingface.co/mistralai

Edit: I'm guessing this is the model referenced: "Mistral "Mixtral" 8x7B 32k model [magnet]" - https://news.ycombinator.com/item?id=38570537

That's it. The community is working on getting it implemented in the popular inference engines. It came out less then 48 hours ago so give it a few days and we should be able to run it and test it at various quantizations depending on how much compute power you have.

There’s an initial version of it already up and running:

https://huggingface.co/DiscoResearch/mixtral-7b-8expert

I am doubtful about these efforts because people are just guessing about how it is supposed to work since Mistral didn't release any code or information yet.

Yes, there are probably some implementation pieces missing. But Disco Research also punches way above their weight on open source LLM stuff, so I would gamble it’s not that far off. They also just released a chat fine-tune of the model just moments ago:

https://huggingface.co/DiscoResearch/DiscoLM-mixtral-8x7b-v2

People had Mixtral 8x7B running on Twitter within 3 hours of the torrent being posted.

It's already been independently benchmarked and finetunes are cooking.

Reply on news.ycombinator.com