Skip Navigation

New open-weight ๐Ÿ‹ DeepSeek V3. 685B MoE. Beats Claude 3.5 Sonnet on Aider coding benchmark

huggingface.co

deepseek-ai/DeepSeek-V3 ยท Hugging Face

Absolutely humongous model. Mixture of 256 experts with 8 activated each time.

Aider leaderboard: The only model above ๐Ÿ‹ 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.

2 comments