MiniMax releases M2.1 coding model
MiniMax's M2.1 extends its open-source M2 line past Python into multi-language and mobile development, and the company says it beats Claude Sonnet 4.5 on multilingual software-engineering benchmarks.
MiniMax announced MiniMax M2.1, an update to its M2 model. The company says M2 was mainly about lowering cost and widening access. M2.1 targets performance on complex real-world tasks, especially coding in languages beyond Python and office work.
MiniMax says it systematically improved the model in Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript and JavaScript. It also says it strengthened native Android and iOS development, which it calls a weakness across the industry. On multilingual software-engineering leaderboards, the company says M2.1 outperforms Claude Sonnet 4.5 and comes close to Claude Opus 4.5. It also reports that M2.1 matches or exceeds Sonnet 4.5 on test-case generation, code performance optimization, code review and instruction following. The post does not list full per-benchmark scores against those models.
MiniMax also built its own benchmark, VIBE (Visual & Interactive Benchmark for Execution), to measure full-stack app building. VIBE has five parts: Web, Simulation, Android, iOS and Backend. M2.1 averaged 88.6. Because MiniMax designed the benchmark and reports the result itself, the score has not been independently verified.
Other claims in the post:
- Shorter outputs: responses and reasoning chains are more concise than M2's, cutting token use and speeding up responses.
- Agent tools: results are consistent in Claude Code, Factory AI's Droid, Cline, Kilo Code, Roo Code and BlackBox, with support for Skill.md, Claude.md/agent.md/cursorrule files and slash commands.
- Interleaved thinking: MiniMax describes the M2 series as one of the first open-source model lines to adopt it systematically.
The post does not give pricing, parameter counts or context length.