MiniMax M2: MiniMax's First Agentic Coding Model

What MiniMax M2 is

MiniMax M2 is the first of MiniMax's agentic coding models, Apache 2.0 licensed and self-hostable, with a 200K token context window — recorded here as 204.8K, the same figure written out — at roughly $0.30 per million input tokens and $1.20 output through third-party providers. The line continues through M2.5 and M2.7, both covered elsewhere on this site.

The workflow MiniMax name

MiniMax describe M2 as engineered for end-to-end developer workflows, listing multi-file edits, coding-run-fix loops and test-validated repairs. They point to Terminal-Bench and Multi-SWE-Bench-style tasks as the evidence, and describe use across terminals, IDEs and CI systems in multiple languages.

External evaluator

MiniMax state they align with Artificial Analysis, which aggregates benchmarks across mathematics, science, instruction following, coding and agentic tool use. The later M2.5 and M2.7 cards lead with MiniMax's own figures.

LMArena

On LMArena's text leaderboard MiniMax M2 has 1344 Elo (rank 230); by category, 1383 Elo in Coding (rank 235), 1353 Elo in Math (rank 216), 1351 Elo in Expert (rank 225) and 1283 Elo in Creative Writing (rank 265). M2.5 has 1391 Elo (rank 179) and M2.7 has 1415 Elo (rank 142).

Architecture and MiniMax’s table

MiniMax-M2 is a mixture-of-experts model with 230 billion total parameters and 10 billion active per token. In MiniMax's own evaluations it records SWE-bench Verified 69.4, Multi-SWE-Bench 36.2, SWE-bench Multilingual 56.5, Terminal-Bench 46.3, ArtifactsBench 66.8, BrowseComp 44, GAIA (text only) 75.7, HLE with tools 31.8, τ²-Bench 77.2 and FinSearchComp-global 65.5. For comparison the same table gives Claude Sonnet 4.5 77.2, 44.3, 68, 50, 61.5, 19.6, 71.2, 24.5, 84.7 and 60.8. Figures marked as taken from other vendors' reports are mixed with MiniMax's own runs.

How the figures were produced

MiniMax state SWE-bench Verified used the R2E-Gym scaffold on OpenHands with a 128K context, 100 maximum steps and no test-time scaling; Multi-SWE-Bench and SWE-bench Multilingual were averaged over eight runs with the claude-code CLI; Terminal-Bench used the official claude-code harness; browsing benchmarks used the WebExplorer agent framework; and τ²-Bench used GPT-4.1 as the user simulator. On the Artificial Analysis methodology, MiniMax report an AA Intelligence score of 61, with AIME25 78, MMLU-Pro 82, GPQA Diamond 78, LiveCodeBench 83 and IFBench 72.

Availability at launch

At launch MiniMax made the API free for a limited time on the MiniMax Open Platform and offered MiniMax Agent, built on M2, free for a limited time. The weights are on Hugging Face, with SGLang and vLLM among the recommended serving frameworks; MiniMax note SGLang had day-0 support.

What MiniMax say the activation size is for

MiniMax describe the 10-billion active parameter budget as shortening the plan, act and verify loop in agent workflows: faster feedback in compile-run-test and browse-retrieve-cite chains, more concurrent runs on the same budget for regression suites, and smaller per-request memory with steadier tail latency. They describe the model planning and executing long toolchains across shell, browser, retrieval and code runners. MiniMax describe the name as a Mini model built for Max coding and agentic workflows. The weights are licensed under Apache 2.0, and the API is documented on the MiniMax platform's text-generation guide.

Official Resources