MiniMax-M2 for agents, Cursor 2.0, and a robot you can preorder

MiniMax opened its M2 model on Monday, and the rest of the week kept pointing the same way: agents are the product now.

MiniMax-M2

M2 is a 230B mixture-of-experts model with roughly 10B parameters active at a time, built for coding and tool-driven agent work, and released with open weights. The ratio is the point. A 10B active footprint means serving it costs far less than a dense model of similar quality, and those are the economics that make an always-on agent affordable. I haven't tested it myself, so I'm treating the vendor's benchmark claims as claims.

Cursor released version 2.0 with Composer, its first in-house coding model, and an interface where several agents work in parallel on separate copies of your code. The company says Composer is fast enough to keep you in flow. I'd say the more telling thing is that a product built on other labs' models decided it needed its own. Moonshot also published Kimi Linear, a hybrid attention design that, per the team, shrinks the memory cache considerably and speeds decoding on long context.

Worlds and robots

BAAI released Emu3.5, a model trained to predict what comes next across text and images, which they pitch as a step toward world models. The robotics headline came from 1X, which opened preorders for its NEO home humanoid at $20,000 or $499 a month, with deliveries planned for 2026. I'd want to know how much of the first units' work is remote operation before I'd believe the autonomy story. It's still the first time a person could actually put down money.

A paper to read slowly

Anthropic's interpretability team reported that Claude models can sometimes detect when a concept has been injected into their internal activations and describe it. The effect is unreliable, and the authors say so. Even so, it hints at how we might one day ask a model about its own reasoning and get a useful answer.

OpenAI also announced a restructuring and released an open safety classifier model, but those launch pages blocked my link check, so I left them out.

What I'd watch next

Moonshot's next move. K2 was strong as an instruct model, and a reasoning version would put pressure on the closed labs.