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.