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