MiniMax-M2

2025-10-27 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax releases MiniMax-M2, a compact 230-billion-parameter MoE with 10 billion active, designed for code and agents, under a modified MIT license, at $0.30 / $1.20 per million tokens.

MiniMax releases MiniMax-M2, a compact 230-billion-parameter MoE with 10 billion active, designed for code and agents, under a modified MIT license, at $0.30 / $1.20 per million tokens. ### Context MiniMax's previous models (Text-01, M1) bet on hybrid attention and very long context, but lagged behind on agentic coding. At the end of October 2025, the closed references for coding agents were Claude Sonnet 4.5 (released at the end of September) and GPT-5, while on the open side GLM-4.6, Kimi K2 and DeepSeek-V3.2 were competing for first place. MiniMax is changing course: fewer active parameters, agent use targeted from the design stage. ### What the model brings - A much more compact model: 10 billion active parameters versus 45.9 billion for M1, for lower cost and latency. - An explicit focus on agents: planning, chained tool calls (terminal, browser, Python interpreter, MCP tools) and integration with Claude Code, Cursor, Cline, Kilo Code or Droid. - A place in the top five of the Artificial Analysis ranking according to the vendor's post, which gives an AA Intelligence index of 61 in the model card. - A price announced at 8% of that of Claude Sonnet 4.5, with a throughput of about 100 tokens per second on the official API. - Open weights under an MIT license with an attribution clause limited to very large services. ### Architecture and specifications - Parameters: 230 billion in total, 10 billion active; 62 layers, 256 experts including 8 active per token (config.json). - Context: 196,608 positions (config.json); not stated in the post. - Modalities: text. Training data and knowledge cutoff: not disclosed. - License: MIT with one modification: if the software or a derivative is used in a commercial product or service with more than 100 million monthly active users or more than 30 million dollars in annual recurring revenue, the mention “MiniMax M2” must appear in the interface. Below these thresholds, commercial use is free. - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M2, with vLLM and SGLang support. ### Cost and access | Item | Value | |---|---| | Official API, input | $0.30 per million tokens | | Official API, output | $1.20 per million tokens | | Free access | until November 7, 2025 (API and MiniMax Agent) | | Self-hosting | hardware not disclosed in the sources consulted | Pricing from the October 27, 2025 post; the current pricing page shows the same amounts for M2, with $0.03 for cache read. ### Benchmarks Scores published by the vendor in the model card; the values marked with an asterisk by the vendor come from third-party sources. | Test | MiniMax-M2 | Claude Sonnet 4.5 | GPT-5 (thinking) | Gemini 2.5 Pro | |---|---|---|---|---| | SWE-bench Verified | 69.4 | 77.2 | 74.9 | 63.8 | | Terminal-Bench | 46.3 | 50 | 43.8 | 25.3 | | BrowseComp | 44 | 19.6 | 54.9 | 9.9 | | GAIA (text) | 75.7 | 71.2 | 76.4 | 60.2 | | tau2-Bench | 77.2 | 84.7 | 80.1 | 59.2 | | AA Intelligence (index) | 61 | 63 | 69 | 60 | Source: MiniMax-M2 model card (Hugging Face); the AA index is measured by the vendor using the Artificial Analysis methodology. ### Versus closed models At the time of its release, the best closed model in the table was GPT-5 (thinking), with an AA index of 69 versus 61 for M2, and Claude Sonnet 4.5 came next at 63. M2 is ahead of Gemini 2.5 Pro on SWE-bench Verified (69.4 versus 63.8) and on Terminal-Bench (46.3 versus 25.3), and ahead of Claude Sonnet 4.5 on BrowseComp (44 versus 19.6). It remains behind on SWE-bench Verified (77.2 for Sonnet 4.5) and on tau2-Bench (84.7). On price, the vendor claims 8% of the pricing of Claude Sonnet 4.5, without detailing the calculation. With 10 billion active parameters and an MIT license, self-hosting a model close to the closed references on agents becomes plausible, which weighs in a sovereignty and cost trade-off. ### Key takeaways M2 shifts MiniMax's positioning toward code and agents with a model that has only 10 billion active parameters and a near-MIT license. It does not surpass GPT-5 or Sonnet 4.5 on the overall index, but it comes close on several agent tests at a much lower price.

Tags : LLM, Open Source, MiniMax, MoE, Agents, Code, MIT modifiée