New MiniMax M2.1 Boosts Multilingual Coding and Agentic Office Automation

MiniMax M2.1 targets real-world, multi-language coding and agentic office automation with improved constraint handling, concise outputs, and stronger Web/App support. Open-source weights and deployment guides available.

New MiniMax M2.1 Boosts Multilingual Coding and Agentic Office Automation

TL;DR

  • MiniMax M2.1 released: available on MiniMax Open Platform (https://platform.minimax.io/docs/guides/text-generation) and open-source model weights on Hugging Face (https://huggingface.co/MiniMaxAI/MiniMax-M2.1)
  • Expanded multilingual coding: Rust, Java, Golang, C++, Kotlin, Objective‑C, TypeScript, JavaScript, and more; stronger WebDev/AppDev with native Android/iOS and improved UI/UX understanding
  • Architectural improvements: composite instruction constraints, Interleaved Thinking for multi-step workflows, concise outputs with reduced token use, and better agent/tool scaffolding generalization
  • Benchmarks (VIBE): aggregate 88.6; VIBE‑Web 91.5; VIBE‑Android 89.7; reported gains on multilingual software engineering leaderboards versus select contemporaries

MiniMax M2.1 arrives with a focus on real-world, multi-language development

MiniMax has released MiniMax M2.1, an update aimed at improving performance on complex, real-world programming tasks and office automation workflows. The model is available on the MiniMax Open Platform and as open model weights on Hugging Face: https://platform.minimax.io/docs/guides/text-generation and https://huggingface.co/MiniMaxAI/MiniMax-M2.1. The original announcement is here: https://www.minimax.io/news/minimax-m21

What’s new in M2.1

MiniMax positions M2.1 as an iteration that shifts priorities from accessibility and cost (the focus of M2) toward stronger usability across multiple programming languages and office scenarios. Key technical emphases include:

  • Expanded multilingual coding capabilities across Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, JavaScript, and others — covering tasks from low-level system work to application-layer development.
  • Stronger WebDev and AppDev support, with explicit improvements for native Android and iOS development and better design and aesthetic understanding for web and app interfaces.
  • Composite instruction constraints and Interleaved Thinking, boosting systematic problem solving and enabling the model to better follow multi-step, constraint-heavy instructions common in office workflows.
  • Concise responses and reduced token use, reported to improve response speed and efficiency for coding and agent-driven continuous workflows.
  • Improved agent/tool scaffolding generalization, with consistent behavior across multiple tool chains and context-management schemes.

Benchmarks and the VIBE suite

M2.1 is reported to deliver noticeable gains over M2 on software engineering leaderboards, particularly in multilingual scenarios where it is said to outperform Claude Sonnet 4.5 and approach Claude Opus 4.5 on some measures. The team introduced a new full-stack benchmark, VIBE (Visual & Interactive Benchmark for Execution), which evaluates Web, Simulation, Android, iOS, and Backend development using an Agent-as-a-Verifier (AaaV) approach to test interactive logic and visual aesthetics in runtime.

  • VIBE aggregate score: 88.6
  • VIBE-Web: 91.5
  • VIBE-Android: 89.7

Additional reported improvements include test-case generation, code performance optimization, code review, instruction following, and long-horizon tool use.

Agentic capabilities and office automation

M2.1 expands the concept of a “Digital Employee.” The model can accept web content in text form and drive mouse and keyboard actions via text-based commands, enabling end-to-end automations across administration, data science, finance, HR, and software development. Demonstrated demos include administrative equipment procurement workflows, project-management triage and status updates, and searching merge requests to identify relevant changes.

Tool and agent integrations listed in the announcement include support for Claude Code, Droid (Factory AI), Cline, Kilo Code, Roo Code, BlackBox, and context-management formats such as Skill.md, Claude.md/agent.md/cursorrule, and Slash Commands.

Showcases

The release highlights a variety of practical demos across stacks and formats, illustrating the model’s full-stack capabilities:

Local deployment and recommended inference stacks

Model weights are open-source on Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-M2.1

Recommended inference frameworks and guides:

Recommended inference parameters: temperature=1.0, top_p=0.95, top_k=40. Tool-calling instructions are available at: https://huggingface.co/MiniMaxAI/MiniMax-M2.1/blob/main/docs/tool_calling_guide.md

How to access

Original announcement: https://www.minimax.io/news/minimax-m21

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community