MiniMax has announced the release of M2.1. The company explained that it has systematically strengthened the processing capabilities for more than eight languages, including Rust, Java, and Golang. Since real-world system development assumes collaboration across multiple languages, the company stated that M2.1 focuses on improvements in this area.

Support for Android and iOS native development, which were previously weaknesses in mobile development, has also been strengthened. Expressive capabilities, such as design understanding for web and apps and 3D simulation, have improved, and the company stated that this will lead "vibe coding" toward practical, production-ready implementation.

M2.1 is the first series of open-source models to systematically introduce Interleaved Thinking. It emphasizes the ability to integrate and execute complex constraints rather than simply focusing on the correctness of the code. The company claims this increases the feasibility of using the model for actual office operations.

Compared to the previous generation M2, responses and chains-of-thought have become more concise. In practical programming experiences, improvements in response speed and reductions in token consumption have been confirmed. It achieves smoother and more efficient operation even in AI coding and agent-driven continuous workflows.

The model demonstrates stable performance across major coding tools and agent frameworks such as Claude Code and Cline. It is also reported to be reliable in supporting context management mechanisms like Skill.md and agent.md. Furthermore, the company stated that it can provide more detailed and structured answers for daily conversation and technical document creation.

Two types of APIs are available: M2.1 and M2.1-lightning. While the results are identical, the latter is specialized for speed. In M2.1, the caching function is enabled automatically, allowing for the optimization of cost and latency without manual configuration. Under the Coding Plan, the lightning version is prioritized based on resource load, enabling users to utilize high-speed inference without a change in price.

Source: MiniMax M2.1: Built for Real-World Complex Tasks, Multi-Language Programming（HN 228pt・81コメント） (HN Search (backfill), 2025-12-26)