Beijing, China – Kunlun Wanwei, a leading Chinese technology company, has announced the release of its latest AI models, the Skywork-OR1 series. These models, including 7B and 32B versions, are specifically designed to excel in mathematical and code reasoning tasks. According to Kunlun Wanwei, the Skywork-OR1 series surpasses the performance of Alibaba’s QwQ-32B model in the same parameter scale and approaches the capabilities of DeepSeek-R1.
The Skywork-OR1 series is fully open-sourced, encompassing model weights, training datasets, and complete code. This comprehensive approach aims to foster collaboration and accelerate advancements in the AI community.
Key Highlights of the Skywork-OR1 Series:
- Superior Performance: The models demonstrate industry-leading reasoning capabilities, particularly in mathematics and code.
- Open Source Initiative: Kunlun Wanwei has made the entire series, including weights, datasets, and code, freely available to developers.
- Focus on Reasoning: The Skywork-OR1 series is designed to overcome limitations in logical understanding and complex task solving.
- Three High-Performance Models:
- Skywork-OR1-Math-7B: A specialized model for mathematics with strong coding abilities.
- Skywork-OR1-7B-Preview: A general-purpose model combining mathematical and coding skills.
- Skywork-OR1-32B-Preview: A flagship version for more complex tasks requiring advanced reasoning.
Technical Details and Training Methodology:
Kunlun Wanwei’s Skywork AI team employed a multi-stage GRPO (presumably a proprietary training method) to optimize data filtering and training processes. This approach significantly improved the models’ reasoning performance. The team has also published a detailed technical blog on Notion, outlining the data processing flow, training methods, and key technical findings, providing a replicable reference for the community.
Commitment to Open Source:
Unlike some other open-source reasoning models that only release model weights, Kunlun Wanwei has adopted a highly transparent open-source strategy. The company believes that this comprehensive approach will contribute to the collective progress of the AI community in reasoning ability research.
Future Developments:
Kunlun Wanwei plans to release official versions of the Skywork-OR1-7B and Skywork-OR1-32B models within two weeks, along with a more detailed technical report sharing their experience and insights in training reasoning models.
Open Source Resources:
- Skywork-OR1 Series: https://github.com/SkyworkAI/Skywork-OR1
- Kunlun Wanwei Tiangong Team Open Source Projects: https://huggingface.co/Skywork01
Conclusion:
Kunlun Wanwei’s release of the Skywork-OR1 series marks a significant step forward in the development of open-source AI models for reasoning. By providing complete access to the models and training data, Kunlun Wanwei is empowering researchers and developers to build upon their work and accelerate innovation in the field of artificial intelligence. The performance of the Skywork-OR1 series, exceeding that of Alibaba’s QwQ-32B, demonstrates the potential of focused research and development in specific areas like mathematics and code reasoning. This open-source initiative is expected to have a significant impact on the AI community, fostering collaboration and driving further advancements in the field.
References:
- Kunlun Wanwei Group Official Announcement
- Skywork-OR1 GitHub Repository: https://github.com/SkyworkAI/Skywork-OR1
- Kunlun Wanwei Tiangong Team Hugging Face Page: https://huggingface.co/Skywork01
Views: 0