apparently they also have a long version with 10 million token context window
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
well, they've had big office in Silicon Valley for a while now, especially ByteDance.
And believe it or not, they work 996 hours!
go Deepseek, some pretty strong claims here. Btw, o1-preview to me is the golden standard in reasoning, but it is so expensive (because it generates so many tokens and it's cost per 1m token is 6x that of gpt-4o). So, it would be good if there is a little competition here.
this will be open weights soon?
I would assume that weights will be available, but my understanding with o1 type is that there are additional reasoning steps. So is just having weights enough to replicate the model?
According to Dylan Patel, Deepseek has > 50k H100s, so pretty decent stash
yes should be, see:
