MiniMax-M2

MiniMax-M2 is a series of AI language models from the company MiniMax, built for writing code and for agents: AI that carries out tasks with tools such as a browser or a terminal. The weights are public, so a company can run the models on its own servers instead of paying per request. Check the licence: M2 allows commercial use, M2.7 (2026) only with MiniMax's permission.

Which models are in the series?

MiniMax released M2 on 27 October 2025, at 8% of the price of Claude Sonnet and twice the speed by its own account. M2.1 followed on 23 December 2025, M2.5 on 12 February 2026 and M2.7 on 18 March 2026. The technical report of 26 May 2026 describes the whole series. On 1 June 2026, MiniMax released a successor, MiniMax M3, with a context window of up to 1 million tokens.

How is it built?

The flagship M2 is a mixture-of-experts model with 229.9 billion parameters, of which 9.8 billion are active per token. Each of its 62 layers has 256 small experts, and a router picks 8 of them for every token, so a token uses about 4% of the weights. Attention is full in every layer, with grouped-query attention: 48 query heads share 8 key-value heads, which keeps the KV cache smaller. MiniMax tested hybrids with sliding-window attention and found them weaker on tasks with more than 32,000 tokens of context. The context window is 192,000 tokens, and pre-training used 29.2 trillion tokens.

According to the report, most of what M2 can do comes from the training after that. Data pipelines build tasks for coding and for knowledge work such as search and office tasks, each with an executable workspace and a reward that can be checked. Forge, MiniMax's reinforcement learning system for long agent runs, keeps the agent, the training and the inference separate, so it can train with any agent scaffold, including black-box agents it only sees through their requests to the model.

What does self-evolution mean for M2.7?

The report says M2.7 "takes an early step toward self-evolution". In MiniMax's own development work, it debugs failed training runs, reads logs, adjusts configurations and improves its own agent scaffold, the software around the model that gives it tools, memory and instructions. MiniMax says M2.7 handles 30 to 50% of the daily iteration workload of its reinforcement learning team. Researchers set the goals and review the major decisions, and the report gives no outside check of these internal figures. The post Can MiniMax-M2.7 improve itself? and the entry on self-evolving agentic models go deeper.

How good is it?

In MiniMax's own measurements, M2.7 scored 56.2 on SWE-bench Pro, 57.0 on Terminal-Bench 2.0 and 52.7 on Multi-SWE-bench, the top score of the models compared on that last test. MiniMax says M2.7 approaches the strongest closed models. In the report's table of benchmarks, it stays below the best of Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.4 and Gemini 3.1 Pro on most coding tests: by 1.5 points on SWE-bench Pro (GPT-5.4: 57.7) and by 18.1 on Terminal-Bench 2.0 (GPT-5.4: 75.1).

Example

Suppose a software agency wants a coding assistant on its own servers, so that client code stays in-house. M2 is published under a modified MIT licence: commercial use is free, and only products with more than 100 million monthly users or 30 million dollars in annual recurring revenue have to show "MiniMax M2" in their interface. In September 2026, the M2.7 licence allows personal and non-commercial use, and commercial use needs written permission from MiniMax. Either way the hardware is substantial: the published M2 weights, mostly stored in 8 bits (quantization), take about 230 GB on Hugging Face.