Today, we’re making MiniMax Code CLI v0.4.12 available to developers worldwide and open-sourcing it under the MIT license.
On FrontierHarness Eval, MiniMax Code achieved a state-of-the-art pass rate, the fastest average task completion time, and the second-lowest token usage.*1
We believe a well-designed harness can unlock more of a model’s capabilities while reducing the time developers spend waiting for results.
By open-sourcing MiniMax Code, we’re enabling developers to inspect how tool calls and permissions are handled and build reliable enterprise applications.
We’re also inviting the community to identify issues and contribute fixes. Through real-world use and community review, we aim to keep strengthening MiniMax Code’s safeguards so developers can use it with greater confidence.
Get started:
Explore the source on GitHub:
github.com/MiniMax-AI/minimax-code
MiniMax Code is evolving rapidly. Together with the community, we want to build a secure, reliable, high-performance harness that continues to unlock more of what models can do.
Intelligence with Everyone.
*1 Evaluation notes
FrontierHarness Eval: GitHub repository
- This third-party evaluation tested MiniMax Code 0.3.2 with Kimi K3 on 30 tasks: 23 passed, 4 failed, and 3 timed out without receiving a score. The pass rate is calculated using all 30 tasks. The providers, checkpointing methods, and time limits for some tasks differ from those used in the published baselines. These results are provided for numerical comparison only and do not represent an official leaderboard ranking.
- Evaluation results: PR #13
- Execution traces: Download ZIP
- Adapter configuration: PR #14
