MiniMax releases M1, an open-weight reasoning model with lightning attention

MiniMax has released MiniMax-M1, the world’s first open-weight large-scale hybrid-attention reasoning model. Built on the MiniMax-Text-01 architecture, it combines a Mixture-of-Experts design with a lightning attention mechanism to enable efficient scaling of test-time compute.

• The model contains 456 billion parameters with 45.9 billion activated per token and supports a native context length of 1 million tokens.
• It consumes approximately 25% of the FLOPs of DeepSeek R1 at a generation length of 100K tokens.
• Training utilized a novel RL algorithm called CISPO, which clips importance sampling weights to stabilize training and improve efficiency.
• Full RL training on 512 H800 GPUs completed in three weeks at a rental cost of $534,700.
• Two versions are released with 40K and 80K thinking budgets, showing performance comparable to or superior to DeepSeek-R1 and Qwen3-235B on complex tasks.

The efficient scaling of test-time compute makes MiniMax-M1 particularly suitable for complex tasks requiring long inputs and extensive reasoning.