MiniMax-M1
World's first open-source large-scale hybrid-attention reasoning model with 1M token context
MiniMax-M1
MiniMax • June 2025
Training Data
Up to mid 2025
MiniMax-M1
June 2025
Parameters
456B (45.9B active)
Training Method
Hybrid MoE with Lightning Attention
Context Window
1,000,000 tokens
Knowledge Cutoff
May 2025
Key Features
1M Context • Lightning Attention • Open Source • Reasoning Focus
Capabilities
Reasoning: Excellent (56% SWE-bench)
Long Context: Outstanding
Efficiency: Excellent
What's New in This Version
First hybrid-attention reasoning model using 25% FLOPs of DeepSeek R1 for 100K sequences
World's first open-source large-scale hybrid-attention reasoning model with 1M token context
What's New in This Version
First hybrid-attention reasoning model using 25% FLOPs of DeepSeek R1 for 100K sequences
Technical Specifications
Key Features
Capabilities
Other MiniMax Models
Explore more models from MiniMax
MiniMax-M2.5
MiniMax's flagship model matching frontier performance at 1/20th the cost with 80.2% SWE-bench Verified
MiniMax-M2.5-Lightning
Ultra-fast variant of M2.5 generating 100 tokens per second at $1/hour continuous operation
MiniMax-M2.1
Lightweight coding-focused model with strong multi-language programming and agentic capabilities
