MiniMax-M2.5
MiniMax's February 2026 230B-A10B coding and agent model under a permissive-style license; M2.7 and M3 have followed.
In short. MiniMax-M2.5 is a large text model for coding, search and office tasks. It has about 230 billion parameters but uses only 10 billion for each token, so it is faster than its size suggests. The license allows commercial use if you keep MiniMax's notice and follow its prohibited-uses list. It needs a multi-GPU server, and MiniMax has since released M2.7 and M3.
Summary
MiniMax released M2.5 on 12 February 2026, about three and a half months after M2, as a model for coding, tool use and search, and office work. It is a text-only mixture of experts: the Hugging Face file holds about 229 billion parameters, with 62 layers, 256 experts of which 8 are used per token, and a context of 196,608 positions. MiniMax's M2 card gives 230 billion in total and 10 billion active, and M2.5 has the same configuration, though its own card does not repeat the figure. MiniMax says it trained M2.5 with reinforcement learning in hundreds of thousands of real-world environments, more than 200,000 of them for coding in over ten languages, using its in-house Forge framework.
MiniMax's card reports 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp with context management. It says M2.5 finishes a SWE-Bench Verified task in 22.8 minutes on average, 37% faster than M2.1 and level with Claude Opus 4.6 at 22.9 minutes, and uses about 20 percent fewer search rounds than M2.1. On other coding agents it reports 79.7 on Droid and 76.1 on OpenCode, and it reports a 59.0% average win rate on its internal GDPval-MM office-work test. All of these results are MiniMax's own, and several of the benchmarks, such as VIBE-Pro, RISE and GDPval-MM, are internal.
The weights are FP8 and total 230 GB. MiniMax's vLLM and SGLang guides say the weights need 220 GB plus 240 GB per million context tokens, and suggest four 96 GB GPUs for a combined cache of 400K tokens or eight 144 GB GPUs for up to 3 million tokens, with a longest single sequence of 196K tokens. By our estimate a 4-bit build needs about 158 GB. In the 30 days to 11 November 2026, Hugging Face counted about 1,134,000 downloads, which may include mirrors and automated pulls. The card says the license is Modified-MIT, while the repository's license file is titled MiniMax Model License, allows commercial use, and asks for a notice and compliance with a prohibited-uses policy. M2.7 and M3 use different terms.
Good for
- Self-hosted coding agents and tool-calling workflows on a multi-GPU server
- Search and research agents that make many tool calls
- Office-document and spreadsheet tasks that MiniMax trained the model for
- Teams that want a commercial-friendly license, unlike the non-commercial license of M2.7
Watch out for
- M2.7 and the M3 line are newer, and their licenses differ, with M2.7 limited to non-commercial use
- The card says Modified-MIT, but the license file is the MiniMax Model License with a notice duty and a prohibited-uses policy, so read it
- It needs about 220 GB for the weights alone, plus roughly 240 GB per million tokens of context
- It reads text only
- The benchmark numbers come from MiniMax, and some of them are internal benchmarks
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the M2.5 (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 |
|---|---|---|---|---|---|
| M2.5 (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
A permissive-style limited license from MiniMax, with a version date of 13 February 2026. It grants a worldwide, royalty-free right to use, copy, distribute, modify and make derivative works of the model, with no revenue cap in the text. You must pass the agreement on, mark changed files, keep MiniMax's attribution notice and follow a Prohibited Uses Policy. The model card and Hugging Face metadata call it Modified-MIT, but the license file in the repository is titled MiniMax Model License.
- Attribution or model name must be shown
- Acceptable-use policy applies
Frequently asked questions
- What is MiniMax-M2.5?
- MiniMax-M2.5 is a large text model for coding, search and office tasks. It has about 230 billion parameters but uses only 10 billion for each token, so it is faster than its size suggests. The license allows commercial use if you keep MiniMax's notice and follow its prohibited-uses list. It needs a multi-GPU server, and MiniMax has since released M2.7 and M3.
- Can I use MiniMax-M2.5 commercially?
- Yes, with conditions. MiniMax-M2.5 is released under the MiniMax Model, which allows commercial use but adds extra rules: Attribution or model name must be shown; Acceptable-use policy applies. Read the license text before you ship a product.
- Is MiniMax-M2.5 open source?
- MiniMax-M2.5 is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax Model adds conditions that the Open Source Initiative definition does not allow.
- Can I run MiniMax-M2.5 on my own computer?
- Only on a machine with a lot of memory. The smallest variant, M2.5 (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-M2.5 released?
- MiniMax-M2.5 was released by MiniMax on February 12, 2026.
- How long a context window does MiniMax-M2.5 have?
- MiniMax-M2.5 supports a context window of up to 192K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- MiniMax-M2.5 model card (Hugging Face) huggingface.co
- MiniMax M2.5 announcement (MiniMax) minimax.io
- MiniMax model release notes (MiniMax Platform) platform.minimax.io
- MiniMax-M2.5 license file, titled MiniMax Model License (Hugging Face) huggingface.co
- MiniMax-M2 model card, for the 230B total and 10B active figures (Hugging Face) huggingface.co
- MiniMax M2.5 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-M2.7 MiniMax | 229B 10B active | 200K | MiniMax M2.7 NC (Non-commercial) | 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 Z.ai (Zhipu AI) | 754B 40B active | 198K | MIT (Permissive) | 520 GB GPU server |