Paper summary: MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention

Problem

Reasoning models generate very long outputs, and standard attention makes long sequences expensive. That’s a problem both for inference and for RL training, which generates many long rollouts.

Method

MiniMax-M1 builds on MiniMax-Text-01 (summarised here in January): a hybrid MoE with lightning attention, 456B total parameters and 45.9B active, with a native context of 1 million tokens (8× DeepSeek-R1, per the abstract). It’s trained with large-scale RL on diverse problems, including sandboxed real-world software-engineering environments. The authors also propose CISPO, an RL algorithm that clips importance-sampling weights rather than token updates.

Key results

From the abstract:

• full RL training on 512 H800 GPUs in three weeks, with a rental cost of $534,700
• two released versions, with 40K and 80K thinking budgets
• comparable or superior to the original DeepSeek-R1 and Qwen3-235B, especially on complex software engineering, tool use and long-context tasks

Limitations

There’s no dedicated limitations section in the extracted text. Results are the developers’ own benchmark comparisons.