Kimi-Dev-72B

Learn from issues and code changes

Moonshot reports about 150 billion tokens of intermediate training from GitHub issues and pull-request commits. It excluded repositories in SWE-bench Verified from this data. The two task roles share a simple process: identify a file, then edit it.Kimi Dev 72B launch and training method

Execution supplies the training signal

The launch description rewards a generated patch when the complete test suite passes inside its container environment. It also describes self-play between candidate fixes and generated tests. These are training and evaluation methods; they do not establish that any new patch is safe.Kimi Dev 72B launch and training method

Use the released conversation format

The downloadable checkpoint includes a Transformers example that uses its tokenizer's chat template before generation. The release is a model checkpoint, not a complete repository automation service.Kimi-Dev-72B model card

Practical guidance. A development tool still needs to collect the relevant files, apply a proposed edit and inspect its effects. Keep human review and project-specific checks in that surrounding workflow.

The base model has separate terms

The Qwen2.5-72B license applies conditions to derivatives and redistribution. It also requires separate authorization for commercial products or services with more than 100 million monthly active users. Moonshot’s MIT label does not establish that those upstream obligations have been removed. Review both sets of terms before deployment or redistribution.Qwen2.5-72B upstream model license