Models — one task each
Small models, in the chat.
A chat has jobs an LLM is the wrong tool for: naming a conversation, and soon others. Each one is a few hundred kilobytes of model, trained for that job alone, running in the browser beside the components — no API call, nothing leaving the page. They are published like the packages: on npm, MIT, with their benchmark beside them.
What makes the list: a job a chat actually has, a model that does it as well as a large one would on that job, and a weight a page can afford. The next candidates are named on the roadmap when they are.