Ollama 0.9.0 adds support for thinking models

Some AI models reason before answering: they produce a kind of rough draft where they try approaches, correct themselves, and arrive at a conclusion. The problem was that this rough draft came back mixed into the real answer, and anyone writing an application had to cut it out by hand, hoping not to cut something useful along with it.

Ollama 0.9.0 fixes exactly that, and nothing else. Reasoning becomes a separate field: the rough draft on one side, the answer on the other. If you want to see it, you read it; if you do not, you ignore it — or switch it off entirely by typing /set nothink in the interactive session.

It is a small, targeted update rather than a rewrite: no new memory management, no performance changes. But it landed exactly when reasoning models were becoming normal on local hardware too, and without that separation every application had to invent its own hand-rolled workaround.