MiniMax-M2.7
MiniMax's March 2026 230B-A10B agent model that helped improve itself, with weights under a non-commercial-only license.
In short. MiniMax-M2.7 is a large text model for coding and long agent tasks. It has about 230 billion parameters but uses only 10 billion for each token. You can download it and run it for personal, research or non-profit use, but any commercial use needs written permission from MiniMax. It needs a multi-GPU server.
Summary
MiniMax announced M2.7 on 18 March 2026 through its API and published the weights on Hugging Face in the second week of April. It has the same size as M2.5: a text-only mixture of experts with about 229 billion parameters, 62 layers, 256 experts of which 8 are used per token, and a configuration listing 204,800 positions, while MiniMax's deployment guide gives 196K tokens as the longest single sequence. MiniMax's M2 card gives 230 billion in total and 10 billion active, and M2.7's configuration is identical. MiniMax calls M2.7 its first model to take a deep part in its own evolution: during development an internal version updated its own memory, built dozens of skills for reinforcement-learning experiments and improved a programming scaffold over more than 100 rounds, which MiniMax says lifted performance on its tests by 30 percent.
MiniMax's card reports 56.22% on SWE-Pro, which it says matches GPT-5.3-Codex, 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, 55.6% on VIBE-Pro, 57.0% on Terminal Bench 2, 39.8% on NL2Repo and a 66.6% medal rate on MLE Bench Lite. For office work it reports a GDPval-AA Elo of 1495, which it says is the highest among open-weight models, and 46.3% on Toolathon. M2.7 also supports native agent teams for multi-agent work. All of these results are MiniMax's own, and its claim that the model cut live production-incident recovery time to under three minutes is one we could not verify.
The weights are FP8 and total 230 GB, with the same memory guidance as M2.5: 220 GB for the weights plus 240 GB per million context tokens, with four 96 GB or eight 144 GB GPUs as examples. By our estimate a 4-bit build needs about 158 GB. The license is the main difference from M2.5. Personal, non-profit, academic and research use is free, but any commercial use, including commercial use of a fine-tuned version, needs MiniMax's prior written authorization and a "Built with MiniMax M2.7" notice. In the 30 days to 11 November 2026, Hugging Face counted about 295,000 downloads.
Good for
- Personal, research and non-profit coding agents, run on a multi-GPU server or a machine with very large memory
- Studying self-improving agent scaffolds and agent-team workflows
- Research and education, which the license allows for free
- Comparing the licensing and quality of M2.5, M2.7 and the M3 family
Watch out for
- Commercial use is prohibited without MiniMax's written authorization, and that includes commercial use of fine-tuned versions and of its APIs
- M2.5 has a more permissive license, and the newer M3 family has its own community license
- It needs about 220 GB for the weights alone, plus roughly 240 GB per million tokens of context
- It reads text only
- All benchmark results are MiniMax's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the M2.7 (FP8) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| M2.7 (FP8) FP8 | 229B 10B active | 526 GB | 263 GB | 158 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
An MIT-style license that covers non-commercial use only. Personal use, including self-hosting for coding and building your own tools, and non-profit, academic and research use are free. Any commercial use needs separate prior written authorization from MiniMax, and where it is authorized you must show "Built with MiniMax M2.7".
- Attribution or model name must be shown
- Acceptable-use policy applies
Frequently asked questions
- What is MiniMax-M2.7?
- MiniMax-M2.7 is a large text model for coding and long agent tasks. It has about 230 billion parameters but uses only 10 billion for each token. You can download it and run it for personal, research or non-profit use, but any commercial use needs written permission from MiniMax. It needs a multi-GPU server.
- Can I use MiniMax-M2.7 commercially?
- Not without separate permission. MiniMax-M2.7 is released under the MiniMax M2.7 NC, which allows research and personal use only.
- Is MiniMax-M2.7 open source?
- MiniMax-M2.7 is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax M2.7 NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run MiniMax-M2.7 on my own computer?
- Only on a machine with a lot of memory. The smallest variant, M2.7 (FP8), needs about 158 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was MiniMax-M2.7 released?
- MiniMax-M2.7 was released by MiniMax on April 12, 2026.
- How long a context window does MiniMax-M2.7 have?
- MiniMax-M2.7 supports a context window of up to 200K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- MiniMax-M2.7 model card (Hugging Face) huggingface.co
- MiniMax M2.7 announcement (MiniMax) minimax.io
- MiniMax model release notes, with the 18 March 2026 API release (MiniMax Platform) platform.minimax.io
- MiniMax-M2.7 Non-Commercial License (Hugging Face) huggingface.co
- MiniMax-M2 model card, for the 230B total and 10B active figures (Hugging Face) huggingface.co
- MiniMax M2.7 vLLM deployment guide (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MiniMax-M2.5 MiniMax | 229B 10B active | 192K | MiniMax Model (Conditional) | 158 GB Big-memory desktop | |
| MiniMax M3 MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 295 GB Big-memory desktop | |
| MiniMax M3.1 Flash MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 277 GB Big-memory desktop | |
| GLM-5.1 Z.ai (Zhipu AI) | 754B 40B active | 198K | MIT (Permissive) | 520 GB GPU server |