What is DeepSWE
DeepSWE is an open-source AI Agent framework by Together.ai and Agentica, trained with reinforcement learning based on the Qwen3-32B model. DeepSWE performs excellently on the SWE-Bench-Verified benchmark, achieving up to 59.0% accuracy with test-time scaling (TTS), and a Pass@1 accuracy of 42.2% without TTS, ranking first among open-source Agent frameworks. The model training data, code, training and evaluation logs, and more are all open-sourced, helping developers learn and improve Agents, and advancing the application and development of reinforcement learning in the field of software engineering.
Main features of DeepSWE
- Code understanding and editing: Understands and modifies existing code to solve specific software problems or implement new features.
- Problem solving: Based on interaction with the environment, solves complex software engineering problems, such as fixing issues on GitHub, implementing new code features, and debugging.
- Automated testing and verification: Runs shell commands for code building and testing, verifies the effectiveness of solutions, and ensures that modified code does not break existing functionality.
- Multi-step reasoning: When solving tasks, performs multi-step reasoning and decision-making, gradually optimizing the solution until the task is finally completed.
Technical principles of DeepSWE
- Reinforcement learning training: Trained entirely from scratch with reinforcement learning (RL), without relying on distillation from a stronger proprietary teacher model or supervised fine-tuning (SFT). Based on interaction with the environment, it learns how to make optimal decisions in complex software engineering tasks.
- rLLM framework: Trained based on the rLLM framework, a system used for post-training language agents, providing efficient data management and training processes, and supporting large-scale reinforcement learning training.
- Sparse outcome reward model: Based on a sparse outcome reward mechanism, a positive reward is given only when the code patch generated by the model passes all tests; otherwise the reward is zero, encouraging the model to learn to generate high-quality solutions.
- Test-time scaling (TTS): During the testing phase, generates multiple trajectories and selects the trajectory that correctly solves the problem, significantly improving the model's accuracy. Combining the advantages of execution-based verifiers and execution-free verifiers achieves a substantial performance improvement.
- Kubernetes support: To address scaling challenges during training, Kubernetes support is integrated into the training environment, enabling elastic scheduling and automatic scaling of containers, ensuring the efficiency and stability of the training process.
DeepSWE project address
- HuggingFace model repository：https://huggingface.co/agentica-org/DeepSWE-Preview
Application scenarios of DeepSWE
- Code understanding and editing: DeepSWE can automatically understand and modify code, quickly locate and fix vulnerabilities, optimize performance bottlenecks, and refactor code structure, significantly improving the overall quality of code.
- Software problem solving: Efficiently handles issues on GitHub, implements new features, and breaks down complex tasks into manageable subtasks, thereby quickly solving various difficult problems in the field of software engineering.
- Automated testing and verification: DeepSWE can automatically generate test cases, execute code builds and automated tests, and perform regression testing, ensuring software stability and reliability and reducing the workload of manual testing.
- Multi-step reasoning and solution optimization: Solve complex problems based on multi-step reasoning, continuously iterate and optimize solutions, accumulate knowledge and experience in the process of solving different problems, and better respond to future challenges.
- Development assistance and collaboration: DeepSWE provides developers with real-time code suggestions and completion prompts, assists with project management and task allocation, promotes collaboration and communication among team members, and improves the efficiency of the entire development team.