Skywork-OR1: new SOTA 32B thinking model with open weight
a year ago
- #AI Models
- #Open Source
- #Machine Learning
- Skywork-OR1 series models released, including Skywork-OR1-Math-7B, Skywork-OR1-32B-Preview, and Skywork-OR1-7B-Preview.
- Open-sourced model weights, training data (coming soon), and code.
- Notion Blog released with detailed training recipes, experimental results, and insights.
- Skywork-OR1-Math-7B optimized for math reasoning, scoring 69.8 on AIME24 and 52.3 on AIME25.
- Skywork-OR1-32B-Preview matches Deepseek-R1 performance on math and coding tasks.
- Skywork-OR1-7B-Preview outperforms similarly sized models in math and coding.
- Final release version available in two weeks.
- Evaluation metrics include AIME24, AIME25, and LiveCodeBench using Avg@K for reliability.
- Detailed performance comparison table provided for various models.
- Docker and Conda setup instructions included for running Skywork-OR1 models.
- Training scripts to be released in 1-2 days.
- Evaluation scripts available for reproducing results on AIME24, AIME25, and LiveCodeBench.
- Technical report to be released soon.
- Models trained on DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B.
- Citation provided for the Skywork Open Reasoner Series.