MiniMaxAI/MiniMax-M2.1

Hugging Face
TEXT GENERATIONPricing:Input $0.3 / Cached $0.06 / Output $1.2Concurrent Unit Cost:4Model Size:229BQuant:FP8Context Size:200kTool Calling:SupportedPublished:Dec 20, 2025License:otherArchitecture:Transformer1.4K Warm

MiniMax-M2.1 is a large language model developed by MiniMaxAI, optimized for advanced agentic capabilities. It excels in coding, tool use, instruction following, and long-horizon planning, particularly in multilingual software development and complex multi-step office workflows. Benchmarks show it outperforms MiniMax-M2 and often matches or exceeds Claude Sonnet 4.5 in software engineering tasks, including a novel VIBE benchmark for full-stack application development. This model is designed to democratize high-performance agentic AI, offering transparency and accessibility for building autonomous applications.

Loading preview...

MiniMax-M2.1: An Open-Source Agentic LLM

MiniMax-M2.1, developed by MiniMaxAI, is an advanced large language model released to the open-source community, focusing on democratizing high-tier agentic capabilities. This model is specifically optimized for robust performance in coding, complex tool use, precise instruction following, and long-horizon planning, aiming to enable the creation of next-generation autonomous applications.

Key Capabilities

  • Superior Software Engineering: Significantly improves over its predecessor, MiniMax-M2, on core software engineering leaderboards like SWE-bench Verified, Multi-SWE-bench, and Terminal-bench 2.0.
  • Multilingual Proficiency: Demonstrates strong performance in multilingual scenarios, often outperforming Claude Sonnet 4.5 and approaching Claude Opus 4.5.
  • Full-Stack Application Development: Achieves an average score of 88.6 on the novel VIBE (Visual & Interactive Benchmark for Execution in Application Development), showcasing robust capabilities in building complete, functional applications across Web, Simulation, Android, iOS, and Backend.
  • Enhanced Agentic Features: Shows steady improvements in long-horizon tool use and comprehensive intelligence metrics, including Toolathlon and BrowseComp.
  • Framework Generalization: Exhibits exceptional framework generalization and robust stability across various coding agent frameworks.

Good for

  • Automating multilingual software development.
  • Executing complex, multi-step office workflows.
  • Building autonomous applications requiring advanced planning and tool use.
  • Developers seeking a transparent and controllable high-performance agentic model for local deployment and custom applications.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p