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

This technical report describes the MiniMax-M1 model, focusing on efficient scaling of test-time compute with Lightning Attention. The work demonstrates significant improvements in software engineering tasks, achieving 64% Pass@1 on SWE-Verified and ranking #2 on MultiSWE and TerminalBench.

Key contributions include the development of a large-scale SWE data synthesis pipeline generating over 36K verifiable tasks from 5K+ sandbox environments, and novel approaches to post-training for software engineering agents.

Download paper here

Recommended citation:

@techreport{minimax2025m1,
  title={MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention},
  author={MiniMax},
  institution={MiniMax},
  year={2025}
}