SpeedyLabsAI/SWE-32B-RL
SpeedyLabsAI/SWE-32B-RL is a 32.8 billion parameter software engineering agent based on Qwen3-32B, specifically trained using pure reinforcement learning on executable repository environments. This model excels at agentic software engineering tasks like bug fixing and patch generation within a tool-use loop. It is designed for thinking-mode operations, producing reasoning alongside code, and is an open reproduction of the SkyRL-Agent recipe.
Loading preview...
SpeedyLabsAI/SWE-32B-RL: A Reinforcement Learning Agent for Software Engineering
SpeedyLabsAI/SWE-32B-RL is a 32.8 billion parameter model built upon Qwen3-32B, uniquely trained with pure reinforcement learning (RL) on executable repository environments. This approach, which avoids supervised fine-tuning or distillation, rewards the model based on whether its generated code patches pass unit tests in a sandbox. It represents an open reproduction of the SkyRL-Agent recipe.
Key Capabilities & Features
- Pure RL Training: Utilizes outcome-reward RL, where the model directly learns from the success or failure of its code edits in real repository environments.
- Agentic Software Engineering: Optimized for tasks such as bug fixing, patch generation, and repository navigation.
- Thinking Model: Designed to output detailed reasoning (
reasoning_content/<think>) alongside its code suggestions. - ReAct Scaffold: Employs a ReAct framework with bash, file-editor, and AST search tools during training, internalizing localization behavior.
- Context Length: Supports a substantial 32,768 token context length.
Intended Use Cases
This model is specifically intended for agentic software engineering within a tool-use loop, compatible with harnesses like OpenHands, Aider, or mini-swe-agent. It is particularly suited for:
- Automated Bug Fixing: Generating patches to resolve failing tests.
- Code Patch Generation: Creating modifications to existing codebases.
- Repository Interaction: Navigating and understanding project structures through tool use.
Limitations
Currently, the model's early checkpoints are comparable to the base Qwen3-32B. It has been trained exclusively on Python-based software engineering tasks (R2E-Gym) and its performance in other domains is untested. Due to its ability to execute commands and edit files, it must be run in a sandboxed environment.