New open-weight 馃悑 DeepSeek V3. 685B MoE. Beats Claude 3.5 Sonnet on Aider coding benchmark
New open-weight 馃悑 DeepSeek V3. 685B MoE. Beats Claude 3.5 Sonnet on Aider coding benchmark
huggingface.co
deepseek-ai/DeepSeek-V3 路 Hugging Face
Absolutely humongous model. Mixture of 256 experts with 8 activated each time.
Aider leaderboard:
The only model above 馃悑 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.