Changes from previous version
Large RL scale-up in real environments versus M2.1; adds M2.5-highspeed variant and much lower cost per token for always-on agents.
Release Summary
MiniMax-M2.5 targets real-world productivity: 80.2% SWE-Bench Verified, strong agentic search and office tasks, and ~37% faster than M2.1 on SWE-Bench Verified runs.
Timeline
February 12, 2026
MiniMax M2.5 released
MiniMax releases M2.5 and M2.5-highspeed for coding, agents, and office workflows.