MiniMax-M1

MiniMax-M1 is the world’s first open-source large-scale hybrid-architecture inference model launched by MiniMax (Shanghai Xiyu Technology), with a core focus on being a productivity-grade large model offering “millions of tokens in context plus ultra-high cost-effectiveness.”

• Ultra-long context support: The industry’s highest input window at 1 million tokens and an output length of 80,000 tokens—far surpassing mainstream models such as DeepSeek R1.
• Innovative hybrid architecture: Based on the proprietary “Lightning Attention Mechanism,” it requires only about 30% of the computational power needed by DeepSeek R1 for deep reasoning on long texts.
• Efficient training via reinforcement learning: Utilizing the self-developed CISPO algorithm, it achieves twice the convergence speed compared to DAPO, completing the RL phase in just three weeks using 512 H800 GPUs.
• Leading performance in productivity scenarios: Outperforms Gemini 2.5 Pro and most closed-source models in tool usage and engineering tasks such as SWE-bench and TAU-bench.

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

Core Capabilities

⚡ Million-token-level efficient context processing: Easily handles entire novels, lengthy codebases, or multi-hour meeting transcripts without losing information or slowing down inference.

🧠 Deep tool collaboration capability: Natively understands API calls, code execution, and environmental feedback, enabling high-success-rate closed-loop operations in agent-based tasks.

🧩 Hybrid-architecture compute optimization: Through dynamic sparse computation and attention compression, it significantly reduces deployment costs while maintaining high performance.

🔓 Open-source and commercially usable: Complete model weights are now available on Hugging Face and GitHub, supporting deployment via mainstream frameworks such as vLLM and Transformers.