OpenHands/CodeScout-14B

Text Generation • 15B • Updated 5 days ago • 82 • 2

Note 🏆 CodeScout-14B — strongest model, SOTA on SWE-Bench Verified/Pro/Lite

OpenHands/CodeScout-4B

Text Generation • 4B • Updated 5 days ago • 88

Note ⚡ CodeScout-4B — outperforms 8× larger Qwen3-32B across all benchmarks

OpenHands/CodeScout-1.7B

Text Generation • 2B • Updated 5 days ago • 83 • 1

Note 🔬 CodeScout-1.7B — post-RL checkpoint, outperforms 8× larger Qwen3-14B

OpenHands/CodeScout-1.7B-RFT

Text Generation • 2B • Updated 5 days ago • 81 • 1

Note 📦 CodeScout-1.7B-RFT — pre-RL (rejection fine-tuned) checkpoint

Note 🗂️ Training rollouts from SWE-Smith environments

Note 📊 Evaluation trajectories on SWE-Bench Verified, Pro, and Lite

Note 🔍 SWE-Smith code search localization targets

Note 🔍 SWE-Gym code search localization targets

Note 🔍 SWE-rebench code search localization targets

Note 🎯 SWE-Bench Verified — localization ground truth

Note 🎯 SWE-Bench Lite — localization ground truth

Note 🎯 SWE-Bench Pro — localization ground truth