MiniMax-M2.7

2026-03-18 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax introduces M2.7, described as its first model to take part in its own evolution, with 56.22% on SWE-Pro, 57.0% on Terminal Bench 2 and weights published under a non-commercial license.

### Context After M2.5 (February 2026), the closed references were Claude Opus 4.6, GPT-5.4 (cited by the vendor) and Gemini 3.1. This time MiniMax highlights a training process in which the model takes part in the research: updating its own memory, creating skills for reinforcement learning experiments and improving its own agent framework. ### What the model brings - Self-evolution: an internal version optimized an agent framework on its own over more than 100 rounds (failure analysis, code modification, evaluation, keep or roll back), with a 30% gain on an internal evaluation set. The vendor states that M2.7 handles 30 to 50% of the workflow of its reinforcement learning team. - Software engineering: 56.22% on SWE-Pro (on par with GPT-5.3-Codex according to the vendor), 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, 55.6% on VIBE-Pro and 57.0% on Terminal Bench 2. - Office work: an ELO score of 1,495 on GDPval-AA, the highest among open models according to the vendor, and 97% skill adherence across more than 40 complex skills. - Native agent teams and dynamic tool search. - Machine learning: 66.6% average medal rate across 22 MLE Bench Lite competitions. ### Architecture and specifications - Parameters: about 229 billion in total (safetensors), 10 billion active like the M2 series (not detailed in the post); 256 experts, 8 of them active, 62 layers. - Context: 204,800 positions (config.json). Modalities: text. - Training data and knowledge cutoff: not disclosed. - License: "Non-Commercial License", according to the LICENSE file. Free for personal use (including self-hosted), for non-profit organizations and for research; any commercial use (paid offering, commercial use of an API, deployment of a fine-tuned model for commercial purposes) requires prior written authorization from MiniMax (api@minimax.io) and the mention "Built with MiniMax M2.7". It is therefore no longer an MIT-type license. - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M2.7, with vLLM, SGLang, Transformers and NVIDIA NIM. ### Cost and access | Item | Value | |---|---| | Official API, input | $0.30 per million tokens | | Official API, output | $1.20 per million tokens | | Cache read / write | $0.06 / $0.375 per million tokens | | Self-hosting | hardware not disclosed in the sources reviewed | Prices from the current pricing page; the highspeed variants charge double. ### Benchmarks The post does not include a legible numerical comparison table across all models. Text figures published by the vendor (post and model card): | Test | MiniMax-M2.7 | Closed reference cited | |---|---|---| | SWE-Pro | 56.22% | GPT-5.3-Codex (tie announced, score not given) | | MLE Bench Lite (average medal rate) | 66.6% | Opus 4.6: 75.7%; GPT-5.4: 71.2%; Gemini 3.1: 66.6% | | VIBE-Pro | 55.6% | close to Opus 4.6 (score not given) | | MM Claw end to end | 62.7% | close to Sonnet 4.6 (score not given) | Source: official post of March 18, 2026 and model card, vendor measurements. ### Versus closed models On MLE Bench Lite, M2.7 scores 66.6%, tied with Gemini 3.1 and behind GPT-5.4 (71.2%) and Opus 4.6 (75.7%), which places it third according to the vendor. On SWE-Pro, the vendor announces a tie with GPT-5.3-Codex without publishing the latter's score, and describes it as close to Opus's top level. On GDPval-AA, the vendor announces an ELO of 1,495, higher than that of GPT-5.3 and the highest among open models. The API price stays at $0.30 and $1.20. The license change is the clearest consequence for the open-or-closed choice: downloading the weights remains possible, but commercial use, including through an internal API within a paid offering, requires prior agreement from MiniMax. ### Key takeaways M2.7 marks a technical turning point (self-improvement of the agent framework) and a legal one (non-commercial license). A company has to plan on obtaining authorization from the vendor, which brings its status closer to that of an available-weights model than of an open model in the usual sense.

Tags : LLM, Open Source, MiniMax, MoE, Agents, Licence non commerciale, Poids ouverts