MiniMax opens a model it says trains itself

For a while now I've been hearing that open source is going to democratize AI and put it in anyone's hands. VentureBeat crowned MiniMax-M2 "the new king of open source models"; MarkTechPost announced that MiniMax had released M2.7, "a self-evolving agent"; and TechNode highlighted that it offered double the speed at 8% of the price of Claude Sonnet. It sounds like the open future is already here, free and in your own home.

MiniMax published M2.7 under a permissive license and presented it as a model able to improve on its own, scoring 56.2% on the SWE-Pro benchmark. The nuance almost nobody mentions: what self-optimizes is the scaffolding around the model —how it organizes itself to solve the task—, not its weights, which don't change by a single digit. It doesn't train itself; it rewrites its own working instructions. The label sells better than the mechanism.

I think the noise is being mistaken for the gift here. Pretending we should all have a data center at home instead of a computer is ignoring the real community that holds open source up. MiniMax is a good model, no doubt: it has found its place in the market and competes very well on price. But it really isn't open source; opening the weights of something almost nobody can run isn't opening anything, it's marketing with a friendly license. They still don't know what they want to be when they grow up.

Sources: VentureBeat · MarkTechPost · Daily Dose of DS (the weights don't change) · TechNode