Cursor released their “Composer” model today. I believe the name is a hat tip to the original name of their agent/multi-file editing mode.

Per the post

Composer is a mixture-of-experts (MoE) language model supporting long-context generation and understanding. It is specialized for software engineering through reinforcement learning (RL) in a diverse range of development environments.

I’ve long been of the opinion that the only way Cursor was going to have staying power as a company was to train their own model. If they hadn’t, my thesis was that they would continue to give all their margin away to the lab(s) with the best coding model(s).

Cursor has a ton of code data. They also see users’ preferences for models.

By training their own model, maybe they can better control their own destiny and not be at the mercy of the labs and their API pricing. The open question is whether this model is good enough to convince their customers to switch to it.