Open R1: Update #4

Source: HuggingFace Date: 2025-03-26 URL: https://huggingface.co/blog/open-r1/update-4

Summary

Model release analysis: DeepSeek-V3-0324 update — same architecture, now MIT licensed. Major benchmark improvements: AIME +19.8 (39.6→59.4), GPQA +9.3 (59.1→68.4), LiveCodeBench +10.0, MMLU-Pro +5.3. Claims parity with GPT-4.5 and outperformance of Claude Sonnet 3.7. Improved function calling, front-end code generation, and Chinese writing quality.

Implications

Thread: open-weights ecosystem health / model release cadence. The MIT license change on DeepSeek-V3-0324 is the most consequential update — it removes the previous commercial use restrictions and makes the model usable for any application. The AIME improvement of +19.8 points in a single update without architecture changes suggests significant post-training improvements. The combination of MIT licensing and frontier-competitive performance continues the pattern established by R1: Chinese open-weight models setting the cost/capability frontier that Western closed models are benchmarked against.