Overview
MiniMax-M2 is a text-generation mixture-of-experts model designed for coding and agent workflows. MiniMax released it on October 27, 2025. The model has 230 billion total parameters, with 10 billion activated per token. (minimax.io)
Capabilities and uses
- Supports end-to-end software development tasks, including multi-file editing, coding, testing and repair.
- Plans and executes multi-step agent workflows involving tools such as shells, browsers, code interpreters and MCP tools. It can also assist with research and general reasoning tasks.
- MiniMax released the complete model weights for local deployment and recommends vLLM or SGLang for serving them. (minimax.io)
Limitations
Results depend on the task, tools and agent framework. Generated code, factual claims and tool actions should be reviewed and tested before use. MiniMax stated at launch that M2 was slightly behind leading overseas models in programming. Local deployment requires suitable computing resources; API availability, limits and current pricing should be checked with the platform. (minimax.io)
API pricing
The following are historical prices announced by MiniMax, not verified current prices. They are separate official quotations, not a currency conversion. Sources were checked on November 17, 2026.
- International announcement, USD: $0.30 per million input tokens and $1.20 per million output tokens. Official international launch announcement. (minimax.io)
- Mainland China announcement, CNY: ¥2.1 per million input tokens and ¥8.4 per million output tokens. Official China launch announcement. (minimax.cn)
- Mainland China, historical cache-read price: ¥0.21 per million cached tokens. The November 7, 2025 announcement specified MiniMax-M2-Stable for its paid API service; this is a version-specific quotation. Official China pricing update. (minimax.cn)
Current prices and availability for this exact model have not been verified; launch-period free offers should not be assumed to remain available.