MiniMax releases M2.1 AI model for multi-language programming versatility – News

Chinese artificial intelligence startup MiniMax today announced the release of M2.1, a significantly enhanced performance for real-world complex tasks and agentic capabilities across more programming languages and office scenarios.

The key highlights of M2.1 include dramatically enhanced programming skills in a multitude of programming languages, including Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript and JavaScript. A jump in capability in aesthetic design capabilities and comprehension for web, Android and iOS user interface development.

M2.1 upgraded systematic problem-solving over M2 and focuses not just on code execution correctness but also on following instructions that include additional or complex guidance. The company said this provides higher usability in real office situations, where many straightforward rules can have many complex nuances.

To deliver this, the company emphasized not just superior coding capabilities, but enhanced dialogue and writing skills. The model excels in everyday conversation, technical documentation and writing alongside the ability to deliver structured responses.

“Our users have come to rely on MiniMax for frontier-grade coding assistance at a fraction of the cost, and early testing shows M2.1 excelling at everything from architecture and orchestration to code reviews and deployment,” said Scott Breitenother, co-founder and chief executive of Kilo Code Inc., an open-source agentic AI coding agent.

MiniMax M2 was released in late October this year. The company stated that M2.1 demonstrated significant improvements in capability over its predecessor, especially in multilingual scenarios. Here, it outperformed Anthropic PBC’s Claude Sonnet 4.5 and approaches Claude Opus 4.5 (the bigger, more complex model).

As part of the evaluations, MiniMax established a new benchmark: VIBE, or Visual and Interactive Benchmark for Execution. The suite encompassed five core capabilities: web, simulation, Android, iOS and backend development. Distinguishing itself from other benchmarks, VIBE is configured in agent-as-a-verifier language. This allows it to assess the interactive logic and visual aesthetics of generated applications.

M2.1 showed what the company called “outstanding performance” on the VIBE benchmark, achieving an average score of 88.6. It particularly excels in the VIBE-Web and VIBE-Android subsets, with scores of 91.5 and 89.7, respectively.

The company also tested the new model against big vendors such as Anthropic, Google LLC, OpenAI Group PBC and DeepSeek across comprehensive industry benchmarks for both coding and knowledge, including MMLU-Pro, Humanity’s Last Exam and Toolathon (for AI agents).

The model showed consistent high performance in agentic tool use, real-world knowledge and complex problem-solving capabilities. It scored 22.0 in HLE w/o tools, a challenging academic benchmark featuring thousands of graduate-level, multi-modal questions across diverse subjects. On MMLU, an equally comprehensive subject-knowledge benchmark, the model scored 88, which is consistently equivalent to or closely behind flagship frontier models.

The model is available as an application programming interface from MiniMax or for download from HuggingFace with open weights (although at the time of writing, the page is not yet available). The company’s flagship service, MiniMax Agent, is built on the new MiniMax-2.1.

Images: Canva, MiniMax