MiniMax M2.7 is now on NetMind.
229.9B params. 9.8B active per token. Near-Opus coding performance. 17x cheaper than Opus 4.6 on input.
$0.30/M input. $1.20/M output. 200K context. One API key. Try it now.
MiniMax just released a technical report for the MiniMax-M2 Series!
The MiniMax-M2 series, a family of Mixture-of-Experts models designed from the ground up for agentic tasks.
The flagship M2 packs 229.9B total parameters but activates only 9.8B per token—thanks to an agent-native RL training system (Forge) and data pipelines built around coding and collaborative workspaces. The latest M2.7 even debugs its own runs and modifies its scaffolding.
The result? M2.7 delivers frontier-tier performance, matching or beating GPT-5.4, Sonnet 4.6, and Opus 4.6 on agentic coding (SWE-bench, Multi-SWE-bench), deep search (Toolathlon, VIBE-Pro), office tasks (MM-ClawBench), and reasoning (MLE-Bench, Artificial Analysis)—all with a fraction of the computational cost.