DeepSeek-V2: A Strong, Economical, and Efficient Moe Language Model

[jasondavies]

https://github.com/deepseek-ai/DeepSeek-V2

[unraveller]

It's claiming to be llama3-70B tier in strength, 3x cheaper, 3-5x faster than it due to only having 21B out of 400B+ activated at any one time. With L3-70B normally costing <$1/Million.

[bearjaws]

It's performance at 21B parameters is very impressive.

I also like using something between 13 and 70B parameters, since it will run on a 32GB MacBook Pro easily.

[thethirdone → bearjaws]

Do note that it has 236 B parameters which makes the weights ~450 GB.