HuggingFace Open R1: reproducible open-source implementation of DeepSeek-R1 reasoning

Category: ML research
Status: developing
Sources: GitHub repository, HN discussion
Summary: HuggingFace's Open R1 project reproduces DeepSeek-R1 chain-of-thought reasoning capabilities with open-source training recipes. Step 1 is complete: the Mixture-of-Thoughts dataset (350,000 verified reasoning traces) and OpenR1-Distill-7B are published, matching DeepSeek-R1 on standard reasoning benchmarks (AIME 2024, MATH-500, GPQA Diamond, LiveCodeBench). Steps 2 and 3, the full RL training pipeline, are ongoing. Model sizes from 0.6B to 70B parameters are supported. The HN thread reached 231 points.
Why it matters: Reproducible, open training recipes for reasoning models let practitioners build domain-specific reasoning models without proprietary data or training infrastructure.