Kimi-Dev-72B
Open 72B coding model (fine-tuned from Qwen2.5-72B) reaching 60.4% on SWE-bench Verified via large-scale RL on real repositories.
Reported the best open-weights SWE-bench Verified score at release by RL-training an issue-resolution model in Docker with reward only when the full test suite passes. A stepping stone before K2's agentic coding.
- Date
- Monday, 16 June 2025
- Lab
- Moonshot AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 60.4% company claim; best open-source at release | company |
MIT license; base model Qwen/Qwen2.5-72B.
Sources
This record was checked against its sources on 6 October 2026. How we check