MiniMax M2 Slashes AI Costs with Open-Source Coding Model

This is not a story about a cheap model pretending to be smart. It’s about a smart model that was built to be cheap on purpose. MiniMax didn’t stumble into efficiency; they engineered it , trading brute force for surgical precision.

That 230-billion-parameter architecture, the one that only wakes up 10 billion at a time? It’s not a compromise. It’s a redefinition of what a frontier model should cost.

Thirty percent of HQ tasks. Eighty percent of new code. These aren’t vanity metrics.

They are proof that a leaner, faster, far more economical machine can own the messy, high-stakes work that drives a real company forward. The math is simple: when you slash cost by 95% without cratering capability, you don’t just save money. You unlock workflows that were never viable before.

You let agents run wild without a CFO panicking. That’s the real promise here. Not a cheaper Claude.

An entirely new calculus for deploying intelligence at scale. MiniMax is eating their own dog food, and their kitchen is running on fumes , with results that make the extravagant players look like they’re burning cash for ego. The era of “good enough and dirt cheap” has arrived.

And it’s already writing commits at HQ.