Kimi-Dev is Moonshot AI's open-source coding model that learned to fix real GitHub issues, finding the bug, patching it and writing the test, by practising inside Docker until whole test suites passed. Here's how it works, what it scored and when it's still worth using.
Top open-model score when released, Moonshot-reported.
A dense model built on Qwen2.5-72B.
One model that both fixes bugs and writes tests.
Weights, code and a peer-reviewed paper are public.
| Spec | Kimi-Dev-72B |
|---|---|
| Released | 17 June 2025 |
| Developer | Moonshot AI (Kimi-Dev team) |
| Model | Kimi-Dev-72B, a 72-billion-parameter dense model built on Qwen2.5-72B |
| Purpose | Resolving real software issues: finding the bug, patching it and writing tests |
| Headline result | 60.4% on SWE-bench Verified (Agentless setup, Moonshot-reported) |
| Training | About 150B tokens of mid-training on real code data, then reinforcement learning in Docker |
| Licence | MIT |
| Where | GitHub (MoonshotAI/Kimi-Dev) and Hugging Face (moonshotai/Kimi-Dev-72B) |
| Paper | “Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents”, ICLR 2026 |
| Hosted API | None. Run it yourself. |
From Moonshot's Kimi-Dev GitHub repository, blog post and paper.
About 150B tokens of real-world GitHub issues, pull requests and commits taught the model how developers fix bugs. Moonshot removed any repository used in SWE-bench Verified.
Worked examples of locating files and editing code gave it a starting point for both roles.
The model patched real repositories inside Docker and was rewarded only when the whole test suite passed, so partial fixes earned nothing.
At answer time it writes several patches and several tests, runs them against each other and submits the patch that passes the most.
Self-play: patches and tests check each other, and the best patch wins.
| Test | Kimi-Dev-72B | What it means |
|---|---|---|
| SWE-bench Verified, Agentless | 60.4% | Best among workflow approaches and open models when released (June 2025) |
| SWE-bench Verified, as an agent | 48.6% pass@1 | After light fine-tuning on 5k public agent runs; on par with Claude 3.5 Sonnet (Oct 2024) |
| SWE-bench-Live and Multilingual | Gains carried over | Paper shows its skills generalise beyond the benchmark it trained for |
Moonshot-reported, from the Kimi-Dev blog and paper. Moonshot said Kimi-Dev beat the much larger DeepSeek-R1 update of May 2025 on SWE-bench Verified at launch.
Moonshot's example serves the full model with vLLM across 8 GPUs. For a single 48 GB GPU, try a community 4-bit (GGUF) version with Ollama or llama.cpp.
Moonshot also offers pre-processed SWE-bench repository data in the GitHub repo to save time when running the benchmark.
| You want to… | Use | Why |
|---|---|---|
| Fix bugs and build features day to day | Kimi Code with K2.8 Preview or K3.1 Pro | Far stronger, 1M context, no hardware to run |
| Research how coding agents learn | Kimi-Dev-72B | Open weights, open code and a detailed paper |
| Self-host a coding model on modest hardware | Kimi-Dev-72B (quantised) | 72B fits on one or two large GPUs when quantised |
| Self-host a frontier coding model | Kimi K2.7 Code or K3 open weights | Newer and much stronger, but need far more hardware |
Compare Kimi's current coding models: K2.8 Preview, K3.1 Pro, K2.7 Code and K3.
Released in June 2025. Kimi's 2026 models score far higher on coding tests, so don't pick Kimi-Dev for quality.
It shines in a two-step locate-then-edit workflow. As a free-roaming agent it needs extra fine-tuning.
A full-precision 72B model needs multiple data-centre GPUs. Quantised versions trade some quality for size.
SWE-bench Verified is mostly Python repositories. Test it on your own languages and codebase.