minimax/minimax-m2

MiniMax M2 is a high-efficiency reasoning large model developed by MiniMax (Shanghai Xiyu Technology), specifically designed for agents and code. Its core positioning is as the next-generation AI-native productivity engine—“high intelligence, low price, and fast response.”

• Agent-specific optimization: Delivers outstanding performance in complex, long-chain tasks, enabling stable and coordinated invocation of Shell, browsers, Python executors, and various MCP tools.
• Top-tier coding capabilities: Performs exceptionally well in mainstream development environments such as Claude Code, Cursor, and Cline, providing a seamless end-to-end programming experience.
• Exceptional cost-effectiveness: Its API price is only 8% of that of Claude Sonnet, with inference speeds nearly doubled. Input costs as low as $0.3 per million tokens.
• Leading overall capabilities: Ranked among the global top five on the Artificial Analysis leaderboard, closely approaching top overseas models in tool usage and deep search capabilities.
• Full-stack open-source and accessible: Model weights have been open-sourced on Hugging Face, supporting deployment via vLLM and SGLang, and offering a free API trial.

───────────────────────────────────────────────────────────────────

Core Capabilities

🧠 Strong planning and execution capabilities: Can autonomously break down complex goals (e.g., “Analyze user feedback and generate a product recommendation report”), invoke tools step by step, and integrate results.

💻 Developer-friendly architecture: Natively understands engineering contexts, supporting full-link coding tasks such as multi-file editing, dependency installation, and error debugging.

🔍 Deep information mining: Combines browser and code execution to provide a one-stop research solution—from web scraping and data cleaning to visualization.

⚡ High-speed, low-consumption inference: Through efficient activation parameter design, it achieves a smooth response experience with TPS ≈100 while maintaining high intelligence.