SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformsSkywork-OR1

Skywork-OR1

Skywork-OR1 (Open Reasoner 1) is a series of math and code reasoning models and accompanying training solutions open-sourced by SkyworkAI. This series uses large-scale rule-based reinforcement learning, combined with carefully designed datasets and training recipes, and includes two general reasoning models, Skywork-OR1-7B and Skywork-OR1-32B. According to official materials, Skywork-OR1-32B outperforms Deepseek-R1 and Qwen3-32B on math tasks (AIME24 and AIME25) and performs comparably on code tasks (LiveCodeBench); Skywork-OR1-7B is competitive in math and coding scenarios among models of the same scale. The project has open-sourced model weights, training data, and code, and provides training and evaluation scripts, supports reproduction via Docker and Conda environments, and supports multi-node training using Ray. The official team also released a technical report (arXiv) and Notion Blog, sharing detailed training recipes, experimental results, and analysis. The project also provides the Skywork-OR1-RL-Data training dataset, as well as evaluation scripts for AIME24/AIME25 and LiveCodeBench. Evaluation uses Avg@K as the primary metric rather than Pass@1, to reduce the impact of randomness.

Open website

About this tool

Skywork-OR1 (Open Reasoner 1) is a series of math and code reasoning models open-sourced by SkyworkAI, along with accompanying training data, code, and training solutions. The project aims to improve math and code reasoning capabilities through large-scale rule-based reinforcement learning, combined with carefully designed datasets and training recipes.

Purpose

The Skywork-OR1 series is mainly oriented toward math and code reasoning tasks. Official materials show that this series of models is evaluated on math tasks (AIME24, AIME25) and code tasks (LiveCodeBench), and uses Avg@K as the primary metric to measure the model's average performance across K independent attempts, reducing the impact of randomness and enhancing the reliability of results.

Main Capabilities

  • Model series: Includes two general reasoning models, Skywork-OR1-7B and Skywork-OR1-32B.
  • Performance: According to official releases, Skywork-OR1-32B outperforms Deepseek-R1 and Qwen3-32B on math tasks (AIME24 and AIME25) and performs comparably on code tasks (LiveCodeBench); Skywork-OR1-7B is competitive in math and coding scenarios among models of the same scale.
  • Open-source content: Model weights, training data (Skywork-OR1-RL-Data), and code have all been open-sourced.
  • Training solution: Provides training scripts and data, and can reproduce the results of the Skywork-OR1 series; supports multi-node training using Ray.
  • Evaluation support: By default, AIME dataset evaluation is provided; to evaluate LiveCodeBench, refer to the evaluation data preparation section to set the corresponding path. Evaluation results are automatically saved to outputs/evalation/pass.csv.
  • Technical report: On May 29, 2025, the Skywork Open Reasoner 1 technical report was published on arXiv, providing the training pipeline, research on and mitigation of the entropy collapse phenomenon, and extensive analysis and ablation studies.

Usage

The project provides two approaches: a Docker environment and a Conda environment. The official team provides training scripts and data to reproduce the results. For training data preparation, a script is provided to download data from Hugging Face and, according to a specific model (such as DeepSeek-R1-Distill-Qwen-{1.5,7,32}B) difficulty level filtering issue. For multi-node training, you need to first use Ray to start and connect all nodes, then launch the training script as usual, and the script will automatically utilize the connected nodes. For evaluation, an evaluation script is provided to reproduce the Skywork-OR1 series results; the evaluation data for AIME24 and AIME25 is already included in the GitHub repository, while LiveCodeBench data needs to be downloaded from Hugging Face.

Use Cases

Suitable for developers and researchers who study, understand, and advance the frontier of open reasoning models, especially those interested in mathematical and code reasoning, reinforcement learning training methods, multi-node training, and evaluation.

Limitations

The materials do not provide specific model inference speed, deployment hardware requirements, license details, and other information, which are not yet available. The specific content of the technical report needs to be consulted on arXiv. Model training is based on DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B, and uses a custom branch of the verl project for training.

Related tools

Browse category