MiniMax-M2.5

2026-02-12 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax releases M2.5, trained by reinforcement learning across hundreds of thousands of environments, with 80.2 on SWE-bench Verified and a speed of 100 tokens per second in the Lightning version.

### Context Three and a half months after M2, M2.5 arrives just as Claude Opus 4.6 has been released and Opus 4.5, Gemini 3 Pro and GPT-5.2 hold the top of the closed-model rankings. On SWE-bench Verified, the vendor claims that its M2 models are improving faster than those of Claude, GPT and Gemini. The stated goal is an agent that can be used continuously without worrying about cost. ### What the model brings - Code: 80.2% on SWE-bench Verified, 51.3% on Multi-SWE-bench, training on more than 200,000 real-world environments and more than ten languages. - Search and tools: 76.3% on BrowseComp with context management, and about 20% fewer search rounds than M2.1. - Speed: 37% less time than M2.1 on SWE-bench Verified (22.8 minutes on average versus 31.3), matching the level of Claude Opus 4.6 (22.9 minutes) according to the vendor. - Office work: average win rate of 59.0% on the internal GDPval-MM evaluation, in collaboration with finance and legal professionals. - Forge training framework and continued use of the CISPO algorithm, with a process reward mechanism for long agent trajectories. ### Architecture and specifications - Parameters: about 229 billion in total, 10 billion active like M2 (parameter count from the safetensors); 256 experts of which 8 active, 62 layers. - Context: 196,608 positions (config.json). Modalities: text. - Training data and knowledge cutoff: not disclosed. - License: MiniMax Model License (version of February 13, 2026), a vendor-specific license that grants a worldwide, free, non-exclusive right to use, copy, modify and distribute, with a prohibited-use policy (military use, minors, disinformation, fully automated decisions with legal effect, discrimination). The Hugging Face metadata designates it as "modified-mit". - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M2.5, with vLLM, SGLang, Transformers and KTransformers. ### Cost and access | Item | Value | |---|---| | M2.5-Lightning (100 tokens/s), input | $0.3 per million tokens | | M2.5-Lightning, output | $2.4 per million tokens | | M2.5 (50 tokens/s) | half the Lightning price according to the post | | Current pricing page (M2.5, legacy) | $0.30 input, $1.20 output, $0.03 cache read | | Self-hosting | hardware not disclosed in the sources consulted | The vendor announces a cost of about $1 per hour of continuous operation at 100 tokens per second. ### Benchmarks Scores from the model card appendix, obtained by the vendor on the basis of the public sets of the Artificial Analysis index. | Test | MiniMax-M2.5 | Claude Opus 4.5 | Claude Opus 4.6 | Gemini 3 Pro | GPT-5.2 (thinking) | |---|---|---|---|---|---| | AIME25 | 86.3 | 91.0 | 95.6 | 96.0 | 98.0 | | GPQA Diamond | 85.2 | 87.0 | 90.0 | 91.0 | 90.0 | | Humanity's Last Exam (no tools) | 19.4 | 28.4 | 30.7 | 37.2 | 31.4 | | SciCode | 44.4 | 50.0 | 52.0 | 56.0 | 52.0 | | AA-LCR | 69.5 | 74.0 | 71.0 | 71.0 | 73.0 | Source: MiniMax-M2.5 model card, appendix, internal vendor measurements. In the text of the post, the vendor also gives SWE-bench Verified with the Droid tool (79.7 versus 78.9 for Opus 4.6) and OpenCode (76.1 versus 75.9). ### Versus closed models At the time of its release, the best closed models were Claude Opus 4.6, Gemini 3 Pro and GPT-5.2. On the knowledge and reasoning tests in the table, M2.5 lags behind: 19.4 versus 30.7 for Opus 4.6 on Humanity's Last Exam and 86.3 versus 95.6 on AIME25. It is on agentic coding that the vendor claims parity: 79.7 versus 78.9 for Opus 4.6 with the Droid tool and 76.1 versus 75.9 with OpenCode. These figures are internal vendor measurements. On price, the vendor claims that the output cost is one tenth to one twentieth of that of Opus, Gemini 3 Pro and GPT-5. For an open-versus-closed trade-off, M2.5 offers self-hosting of a model close to the closed references in coding, under a license to be read before any commercial use. ### Key takeaways M2.5 is the milestone where an open model with 10 billion active parameters is presented as on par with the best closed models in agentic coding, at a much lower price. The gains are less clear-cut on scientific reasoning and knowledge.

Tags : LLM, Open Source, MiniMax, MoE, Agents, Code, MiniMax Model License