MiniMax-M2.1

2025-12-23 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax has released M2.1, an update to M2 focused on multilingual programming (Rust, Java, Go, C++, Kotlin), mobile development and interleaved thinking, with 74.0 on SWE-bench Verified.

### Context According to MiniMax, M2 had mainly addressed the question of cost and accessibility. In late December 2025, the closed references were Claude Opus 4.5 (80.9 on SWE-bench Verified), GPT-5.2 and Gemini 3 Pro, and open models were progressing fast with DeepSeek V3.2. M2.1 aims for robustness in real-world conditions rather than a new size tier. ### What the model brings - Multilingual programming: targeted gains on Rust, Java, Go, C++, Kotlin, Objective-C, TypeScript and JavaScript. SWE-bench Multilingual goes from 56.5 (M2) to 72.5. - Web and mobile development: stronger native Android and iOS development, with an average VIBE score of 88.6 versus 67.5 for M2. - Interleaved Thinking: the vendor presents M2.1 as one of the first open families to introduce it systematically, to better handle composite instructions. - More concise answers and shorter reasoning chains, with fewer tokens consumed. - Compatibility with the tools Claude Code, Droid, Cline, Kilo Code, Roo Code and BlackBox, and with the Skill.md, Claude.md and slash commands context mechanisms. ### Architecture and specifications - Parameters: about 229 billion in total (safetensors parameter count), 10 billion active like M2; 256 experts of which 8 active. - Context: 196,608 positions (config.json). - Modalities: text. Training data and knowledge cutoff: not disclosed. - License: modified MIT. Unlike M2, the repository clause requires displaying "MiniMax M2.1" in the interface for any commercial product or service, with no user or revenue threshold. - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M2.1 ### Cost and access | Item | Value | |---|---| | Official API, input | $0.30 per million tokens | | Official API, output | $1.20 per million tokens | | Cache read | $0.03 per million tokens | | Self-hosting | hardware not disclosed in the sources consulted | Pricing from the current pricing page, classified as "legacy"; the launch price was not recorded separately. Fast variants charge double. ### Benchmarks Scores published by the vendor in the M2.1 model card. | Test | MiniMax-M2.1 | Claude Opus 4.5 | GPT-5.2 (thinking) | Gemini 3 Pro | |---|---|---|---|---| | SWE-bench Verified | 74.0 | 80.9 | 80.0 | 78.0 | | Multi-SWE-bench | 49.4 | 50.0 | not measured | 42.7 | | SWE-bench Multilingual | 72.5 | 77.5 | 72.0 | 65.0 | | Terminal-bench 2.0 | 47.9 | 57.8 | 54.0 | 54.2 | | BrowseComp (context management) | 62.0 | 57.8 | 70.0 | 59.2 | | Toolathlon | 43.5 | 43.5 | 41.7 | 36.4 | Source: MiniMax-M2.1 model card (Hugging Face), vendor measurements. ### Versus closed models At the time of its release, the best closed model in the table on SWE-bench Verified was Claude Opus 4.5 (80.9), 6.9 points above M2.1 (74.0). The gap is smaller on Multi-SWE-bench (49.4 versus 50.0) and SWE-bench Multilingual (72.5 versus 77.5 for Opus 4.5, while narrowly beating Gemini 3 Pro at 65.0 and GPT-5.2 at 72.0). M2.1 matches Claude Opus 4.5 on Toolathlon (43.5) and exceeds it on BrowseComp with context management (62.0 versus 57.8), but remains behind on Terminal-bench 2.0 (47.9 versus 57.8). The price remains that of M2 ($0.30 and $1.20), far below the pricing of the closed models cited, whose pricing is not included here. For an open or closed choice, M2.1 becomes a credible option for multilingual code, with an attribution clause to comply with for any commercial use. ### Key takeaways M2.1 consolidates M2 on multilingual code, mobile and tool use, with a reduced gap against the closed models of December 2025. The license changes in nature compared with M2: attribution becomes mandatory for any commercial use.

Tags : LLM, Open Source, MiniMax, MoE, Code, Agents, MIT modifiée