Definition
Plain language
A mixture-of-experts model that uses only a small fraction of its parameters per token but competes with frontier systems on agent tasks.
As stated in the literature
MiniMax's ~230B-parameter fine-grained mixture-of-experts model with ~10B active per token, optimized for agentic workloads via verifiable-reward data pipelines and the Forge training system.
Also called: MiniMax M2, M2, M2.5, M2.7
Why it matters: It shows that sparse routing plus verifiable-reward training can match much larger dense models on real agent workloads.
For example, the model can route each token through just ~10B of its ~230B parameters, getting frontier-level agent performance at a fraction of dense-model cost.
Heard on the show
“Take Apple's M1, M2, M2 Pro, M4.”Episode 158 — How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave