MiniMax releases MiniMax-M2 for coding and agents

Background

MiniMax is a Chinese AI company. In late October 2025 it released MiniMax-M2, which it described as 'a Mini model built for Max coding & agentic workflows' ^[1]. Agentic workflows are tasks in which an AI system works through several steps on its own, calling outside tools and using their results to decide what to do next.

What happened

MiniMax released M2 in late October 2025. The model's weights, the trained numerical parameters that let other people download and run it, were posted to Hugging Face on 22 October 2025 ^[2][3]. Hugging Face is a platform where researchers and developers share models and datasets.

Key details

M2 is a mixture-of-experts (MoE) model. Such a model is split into many specialised sub-networks, called experts, and only some of them are activated for each input. M2 has 230 billion parameters in total and 10 billion active at a time ^[1], so each step uses far fewer parameters than the full size suggests.

M2 is also an interleaved thinking model. It reasons inside `<think>` blocks between its actions. MiniMax said that this thinking content must stay in the conversation history to get the best performance, so developers building on the model need to send it back on later turns ^[1].

MiniMax's published scores put M2 at 69.4 on SWE-bench Verified, a test that asks models to fix real issues from open-source software projects. Claude Sonnet 4.5 scored 77.2 on the same test ^[1]. M2 scored 46.3 on Terminal-Bench, which tests tasks carried out in a command-line terminal, and 44 on BrowseComp, a test of finding hard-to-locate information on the web ^[1].

On the Artificial Analysis Intelligence Index, which combines several evaluations into one score, M2 scored 61. MiniMax said that placed it first among open-source models, while Claude Sonnet 4.5 scored 63 on the same index ^[1].

Impact

The API, the interface that lets other software send requests to the model, was free for a limited time on the MiniMax Open Platform ^[1]. The launch gave developers a free way to try an open model aimed at coding and agents, and it started the M2 line that MiniMax kept developing afterwards.