New open-weight 🐋 DeepSeek V3. 685B MoE. Beats Claude 3.5 Sonnet on Aider coding benchmark
Absolutely humongous model. Mixture of 256 experts with 8 activated each time.
Aider leaderboard:
The only model above 🐋 v3 here is OpenAI o1. DeepSeek is known to make amazing models and Aider rotates their benchmark over time, so it is unlikely that this is a train-on-benchmark situation.
Some more benchmarks: on Reddit.
https://huggingface.co/deepseek-ai/DeepSeek-V3Open linkView original on mander.xyz
Someone managed to run it on a cluster of Mac Minis lol https://blog.exolabs.net/day-2/