MiniMax M1

MiniMax’s latest open-weight foundation model released in June 2025 as part of MiniMaxWeek.

Key Specifications

• Release Date: June 2025
• License: Open-weight (free download on GitHub)
• Architecture: Hybrid Mixture-of-Experts with Lightning Attention
• Compute Efficiency: 75% reduction vs DeepSeek R1
• Training Method: CISPO (halves training steps needed)

Technical Innovations

Lightning Attention

Proprietary attention mechanism enabling efficient long-context processing with reduced memory and compute requirements.

CISPO Training

Novel training methodology that achieves equivalent model performance in half the training steps compared to conventional approaches.

Efficiency Gains

• 75% compute reduction compared to DeepSeek R1
• Optimized for both inference speed and training cost
• Suitable for resource-constrained deployments

Availability

• Free download on GitHub
• API access through MiniMax platform
• Enterprise deployment options

Competitive Position

Positioned against:

• DeepSeek R1
• Llama 3.x series
• Qwen models
• Mistral Large

See Also

• MiniMax-Text-01 - Earlier text model
• MiniMax-VL-01 - Vision-language model
• MiniMax Speech-02 - Text-to-speech model