Skywork-OR1-32B-GPTQ-Int4

🤔 Skywork-OR1 (Open Reasoner 1)

</div> <div> <br> <div align="center">

</div>

🔥 News

• May 13, 2025: We release the final version of Skywork-OR1 (Open Reasoner 1) series of models, including Skywork-OR1-32B and Skywork-OR1-7B. We open-source

  • 🤗 Model weights: Skywork-OR1-32B, Skywork-OR1-7B
  • 🤗 Training data: Skywork-OR1-RL-Data
  • 🧑‍💻 Code: Skywork-OR1

  The complete technical report is coming soon. See our Notion Blog for training recipes and preliminary experimental results. More analysis and insights will be included in the final technical report, which is dedicated to helping the community better research, understand, and push the frontier of open reasoning models.

📖 Overview

<div align="center"> <img src="./assets/figure_1.jpeg" width="100%"/>

_{The AIME24 and AIME25 scores versus training steps of Skywork-OR1-32B in our training pipeline.}

</div>

The Skywork-OR1 (Open Reasoner 1) model series consists of powerful math and code reasoning models trained using large-scale rule-based reinforcement learning with carefully designed datasets and training recipes. This series includes two general-purpose reasoning modelsl, Skywork-OR1-7B and Skywork-OR1-32B.

• Skywork-OR1-32B outperforms Deepseek-R1 and Qwen3-32B on math tasks (AIME24 and AIME25) and delivers comparable performance on coding tasks (LiveCodeBench).
• Skywork-OR1-7B exhibits competitive performance compared to similarly sized models in both math and coding scenarios.

📊 Evaluation

<div align="center"> <img src="./assets/figure_3.jpeg" width="75%"/> <img src="./assets/figure_2.jpeg" width="75%"/> </div> </div>

We evaluate our models on AIME24, AIME25, and LiveCodeBench. Instead of using Pass@1, which is common in prior work, we introduce Avg@K as the primary metric. This metric robustly measures a model's average performance across K independent attempts, reducing the impact of randomness and enhancing the reliability of the results. We believe that Avg@K provides a better reflection of a model's stability and reasoning consistency.

We include the detailed results in the following table.

See our github repo here for the detailed reproduction of the evaluation results.

⚙️ Training Recipe

We offer a brief overview of our data and training pipeline below. For more details, please refer to our Notion Blog here.

Data

• We select, clean, and curate a dataset of 110K verifiable, challenging, and diverse math problems and 14K coding questions from open-source datasets.
• We perform model-aware difficulty estimation for each problem and model and conduct rigorous quality assessment prior to training to ensure training efficiency and effectiveness.

Training

We develop a customized version of GRPO that leverages both data-wise and training-wise improvements:

• We perform both offline and online difficulty-based filtering and rejection sampling to improve training efficiency.
• We incorporate a multi-stage training pipeline coupled with adaptive entropy control and other techniques to enhance exploration and stability.

📄 Technical Report

Our technical report will be released soon. Stay tuned!

🙏 Acknowledgements

• Both of our models are trained on top of DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B.
• Both models are trained using a custom fork of the wonderful verl project.

📚 Citation

We will update the citation once the technical report is released. In the meantime, please cite the following:

@misc{skywork-or1-2025,
  title={Skywork Open Reasoner Series},
  author = {He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and Liu, Yang and Zhou, Yahui},
  howpublished={\url{https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680}},
  note={Notion Blog},
  year={2025}
}