MiniMax M1

MiniMax’s first open-source large-scale hybrid-attention reasoning model. Two variants: 80K and 40K reasoning-output. 1M token context window, 80K reasoning output (8× DeepSeek R1 at the time). Open weights on Hugging Face.

Why it matters

The “lightning attention” hybrid architecture is the technical moat — efficient long-context processing that lets the model use its full 1M window without the throughput collapse that hits standard attention at long sequences. Anchors a model family that continued into M2.7.

Status in the wiki

• M1 is the open-weights generation covered here (June 2025)
• M2.7 is the closed-source successor at 20× lower cost than Opus 4.6 (covered in summary-damian-malliaros-minimax-m27)

Sources

• MiniMax Agent walkthrough

See Also

• minimax-m27 — successor
• deepseek, qwen, kimi-k2 — sibling open-weights models
• vllm — recommended runtime for serving M1 with high throughput