MiniMax launches M3 model with coding, million-token context, and native multimodal features

MiniMax has officially launched the M3 model, its flagship that combines top-notch coding and agent abilities, a million-token context window, and native multimodal features. The model excels in coding and agent evaluations, handling tasks like breaking down problems, calling tools, and multi-step reasoning to produce ready-to-use code.

Using its own MSA architecture, the M3 API supports up to 1M tokens context, with at least 512K tokens guaranteed. In the BrowseComp agent test, M3 scored 83.5, beating Opus 4.7's 79.3. It also ran for nearly 12 hours to independently reproduce an ICLR 2025 paper, making 18 commits and 23 charts.

Available in standard and high-speed versions with identical results, M3 supports auto-caching. It will soon be open-sourced on HuggingFace and GitHub, allowing private deployment and fine-tuning. Pricing for contexts up to 512K tokens is 2.1 yuan per million tokens for input and 8.4 yuan for output, with a 50% discount for the first 7 days.