Skywork-OR1
Open RL-trained reasoners for math and code: the 32B beats DeepSeek-R1 on AIME, with weights, data, and code all public.
About
Skywork-OR1 (Open Reasoner 1), released by Kunlun's Skywork AI on May 13, 2025, is a from-scratch-style reinforcement-learning project built on the DeepSeek-R1-Distill series. The team applied large-scale rule-based RL with a carefully designed data recipe and published an unusually honest technical report about fighting entropy collapse during training. The result: average accuracy across AIME24, AIME25, and LiveCodeBench rose from 57.8% to 72.8% for the 32B model, which the authors report surpasses both DeepSeek-R1 and Qwen3-32B on the AIME math benchmarks while matching them on LiveCodeBench coding.
What sets OR1 apart is the open-science packaging: model weights (32B and 7B, plus a math-specialized 7B), the RL training dataset, the training code, and the full technical report are all public. Context length is 131K tokens. It's a single-generation release with no successor announced, and the licensing is listed as MIT on the Hugging Face model card — worth a glance at the card itself before commercial use, since I couldn't re-verify the license text in this pass. For open math reasoning at 32B, it's one of the strongest options available.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- AIME24/AIME25: Skywork-OR1-32B surpasses DeepSeek-R1 and Qwen3-32B (Skywork-reported, arXiv 2505.22312)
- LiveCodeBench: comparable to DeepSeek-R1 and Qwen3-32B (Skywork-reported)
- Average accuracy across AIME24/AIME25/LiveCodeBench: 57.8% → 72.8% after RL (Skywork-reported)
- No independent benchmark replication verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Skywork release May 13, 2025 with final 32B/7B weights; technical report arXiv 2505.22312 (May 29, 2025) with full RL pipeline, ablations, and entropy-collapse analysis; weights + RL dataset + training code all published. Open weights. Release facts corroborated across the Skywork GitHub, HF, and arXiv. Strengths: exemplary open-science packaging, strong provider-reported math gains, honest technical reporting. Docked for: (1) all benchmark figures are Skywork-reported with no independent replication found; (2) single release with no successor or maintenance signal; (3) MIT license listed on HF model card but not re-verified in this pass. License: MIT (per HF model card metadata).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Skywork AI (Kunlun) released final 32B/7B on 2025-05-13; RL on DeepSeek-R1-Distill base
- · Skywork-reported: 32B surpasses DeepSeek-R1 and Qwen3-32B on AIME24/AIME25; avg 57.8% -> 72.8%
- · Technical report arXiv 2505.22312 with ablations and entropy-collapse analysis; weights + RL data + code public
- · Single release, no successor; 131K context per community deployment docs