DeepSWE: Training a Fully Open-sourced, State-of-the-Art Coding Agent by Scaling RL
FreeOpen-source SOTA coding agent trained with reinforcement learning
FreeFree tier
Inputs: code
About DeepSWE: Training a Fully Open-sourced, State-of-the-Art Coding Agent by Scaling RL
DeepSWE-Preview is a state-of-the-art, fully open-source coding agent developed through a collaboration between Agentica and Together AI. Trained entirely with reinforcement learning (RL) starting from Qwen3-32B, it achieves 59% on SWE-Bench-Verified with test-time scaling—setting a new SOTA for open-weight coding agents (42.2% Pass@1, 71.0% Pass@16). The model was trained over six days on 64 H100 GPUs using 4,500 real-world software engineering tasks from the R2E-Gym environment. All components—model, dataset, code, training logs, and evaluation logs—are open-sourced to help the community advance RL-based agentic systems.
Key Features
Reasoning-enabled coding agent trained from Qwen3-32B using pure reinforcement learning
Achieves 59% on SWE-Bench-Verified with test-time scaling (42.2% Pass@1, 71.0% Pass@16)
Trained on 4,500 real-world SWE tasks over six days on 64 H100 GPUs
Fully open-sourced: model weights, dataset, training code, and evaluation logs
Built using Agentica's rLLM framework for post-training language agents
Handles multi-step agentic tasks: navigating codebases, editing files, running shell commands, and iteratively verifying solutions
State-of-the-art performance among open-weight coding agents
Pros & Cons
Pros
- State-of-the-art performance among open-weight coding agents on SWE-Bench-Verified
- Fully open-source — model, dataset, training code, and logs are publicly available
- Trained using reinforcement learning without supervised fine-tuning, demonstrating RL scalability for agents
- Test-time scaling further improves accuracy (Pass@16 significantly higher than Pass@1)
- Built on a strong foundation (Qwen3-32B) with transparent training recipe
Cons
- Training requires substantial compute (6 days on 64 H100 GPUs), not easily reproducible without significant resources
- Currently a preview release; further improvements and stable version may be needed
- Performance depends on test-time scaling (Pass@16 is much better than Pass@1), which adds inference cost
- Benchmarked primarily on SWE-Bench-Verified; generalizability to other software engineering tasks may vary
Best For
Autonomous software engineering: resolving GitHub issues, implementing new code features, and debuggingNavigating and contextualizing large codebases for code generation and repairRunning build and test commands to iteratively refine code changesServing as a research baseline for scaling reinforcement learning in agentic tasks
FAQ
What is DeepSWE?
DeepSWE-Preview is a reasoning-enabled coding agent trained entirely with reinforcement learning starting from Qwen3-32B. It is a fully open-source model designed for autonomous software engineering tasks.
How does DeepSWE perform on benchmarks?
It achieves 59% on SWE-Bench-Verified with test-time scaling, with a Pass@1 of 42.2% and Pass@16 of 71.0%, setting a new state-of-the-art among open-weight coding agents.
Is DeepSWE open-source?
Yes. The model weights, training dataset (4,500 real-world SWE tasks), training code (using rLLM framework), and evaluation logs are all publicly released.
What hardware was used to train DeepSWE?
Training was performed over six days on 64 H100 GPUs.