MiniMax-M2

What MiniMax-M2 is

MiniMax-M2 is a large language model released by MiniMax in late October 2025, with its weights posted publicly on Hugging Face on 22 October 2025.^[2][3] It is a mixture-of-experts (MoE) model with 230 billion total parameters, of which 10 billion are active at any one time.^[1] MiniMax describes it as "a Mini model built for Max coding & agentic workflows", framing the design goal around software engineering tasks and multi-step tool use.^[1]

The mixture-of-experts arrangement is what makes the size numbers workable. Instead of running every parameter for every token, a routing step sends each token to a small subset of specialist sub-networks. The model can therefore hold a large pool of parameters on disk while only paying the compute cost of a much smaller one per token — here, roughly one twenty-third of the total.

Thinking, and how to run it

MiniMax-M2 is an interleaved thinking model. It writes out reasoning in a `<think>` block before producing its answer, and can reason again in further blocks as a conversation or a tool-using session proceeds. MiniMax's guidance is that the `<think>` content must be kept in the conversation history for the model to perform best.^[1] In practice that means an application built on M2 should pass earlier reasoning turns back to the model rather than stripping them out to save context.

The model card also gives sampling settings: temperature 1.0, top_p 0.95 and top_k 40.^[1]

Results

On the benchmarks MiniMax reports, M2 scores 69.4 on SWE-bench Verified, a test of fixing real issues in open-source repositories, with Claude Sonnet 4.5 at 77.2 on the same test. It scores 46.3 on Terminal-Bench, which measures work in a command-line environment, and 44 on BrowseComp, a browsing-and-research benchmark.^[1]

On the Artificial Analysis Intelligence Index, an aggregate score compiled from a battery of evaluations, M2 scored 61. MiniMax said that figure ranked first among open-source models; Claude Sonnet 4.5 scored 63 on the same index.^[1]

The model card also sets M2 against other models on its own table. GPT-5 in thinking mode scores 74.9 on SWE-bench Verified and 54.9 on BrowseComp, against 69.4 and 44 for MiniMax-M2, and scores 69 on the Artificial Analysis Intelligence Index, where Claude Sonnet 4 scores 57.^[1][1] Read together, the numbers place M2 close to, but behind, the strongest proprietary models of the period on coding, while its headline claim is leadership within the open-weights field.

Access and licensing

For a limited time, the API was free on the MiniMax Open Platform, which let developers try the model without a paid account.^[1] The published weights made local deployment possible as well.

The licensing signals are mixed. The model page links an MIT license, while the metadata tag on the same page reads modified-MIT.^[1] Anyone planning to build on M2 commercially would need to read the actual license text rather than the summary, since the two labels imply different obligations.

Where it sits

M2 arrived as part of a rapid sequence: MiniMax released M2, M2.1 and M2.5 within about three and a half months.^[2] That cadence matters for how the model should be read — M2 is a snapshot of a line that moved quickly, and its benchmark numbers describe a specific release rather than a settled capability level.

Its significance lies in the combination rather than any single number: an open-weights model with a large parameter pool but a small active footprint, tuned explicitly for coding and agentic work, released with a permissive-looking license and a free API window. For teams that could not send code to a closed API, or that wanted to run a coding-oriented model on their own hardware, that was a meaningful option to have.