DeepSeek R3 Achieves State-of-the-Art Reasoning on Math and Code Benchmarks

DeepSeek has released R3, the newest member of its reasoning-optimized model family, claiming top-of-leaderboard results across every major math and coding benchmark. On the 2026 AIME competition set, R3 scored 94.2%, surpassing both OpenAI's GPT-5 and Google's Gemini 2.5 Pro. The model uses a mixture-of-experts architecture with 671B total parameters, activating 37B per forward pass.

Crucially, DeepSeek has released R3 as fully open-weights under an MIT-equivalent license, continuing its pattern of open releases that has repeatedly disrupted the AI industry. Researchers have already reported successful quantized 4-bit inference on a single H100 GPU at roughly 35 tokens per second — a remarkable throughput for a model at this scale.

The release is accompanied by a detailed technical report that describes a novel reinforcement learning approach dubbed 'recursive chain-of-thought distillation', which teaches the model to self-correct its reasoning traces during RLHF. Industry observers expect this technique to be rapidly replicated by labs worldwide within weeks.