OpenHands/CodeScout-4B
OpenHands/CodeScout-4B is a 4 billion parameter model from the CodeScout family, developed by OpenHands, specializing in repository-level code localization. Trained with GSPO reinforcement learning from Qwen3-4B-Instruct-2507, it excels at identifying relevant files, classes, and functions within a codebase using only standard Unix terminal commands. This model demonstrates superior efficiency-performance trade-offs, outperforming significantly larger models like Qwen3-32B and GPT-5 with RepoNavigator on SWE-Bench code localization benchmarks.
Loading preview...
CodeScout-4B: An Efficient Code Localization Agent
CodeScout-4B is a 4 billion parameter model developed by OpenHands, part of the CodeScout family of open-source, RL-trained code search agents. It is specifically designed for repository-level code localization, identifying relevant files, classes, and functions within a codebase based on a given GitHub issue description.
Key Capabilities & Performance
- Superior Efficiency: Consistently outperforms 8x larger models like Qwen3-32B across all benchmarks.
- High Accuracy: Surpasses RepoNavigator-14B by 2-10% in file F1 and 8-11% in function F1. It also exceeds GPT-5 with RepoNavigator by 9% in file F1 and 5% in function F1 on SWE-Bench Verified.
- Terminal-Based Operation: Achieves state-of-the-art results using only a standard Unix terminal and basic commands (e.g.,
rg,find,grep,ls), without requiring static analysis or language-specific tooling. - Reinforcement Learning: Trained directly from
Qwen3-4B-Instruct-2507using GSPO (Group Sequence Policy Optimization) with multi-level F1 rewards on the SWE-Smith dataset.
Intended Use & Limitations
CodeScout-4B is ideal for integration as a localization subagent within larger coding agent pipelines. It is specifically trained and evaluated on Python repositories for code localization, not code editing or issue resolution. Optimal performance requires the OpenHands-Bash scaffold.