Ollama v0.34.3 has arrived, and with it a feature that would have sounded philosophical three years ago: you can now ask a model how hard it is thinking, and tell it to do more or less of that. The machines, for their part, are complying.
What happened
The GET /api/show endpoint now returns each model's available thinking controls alongside its defaults. For a model like Gemma 4, that means a simple boolean — thinking on or off — with the default set to on, which is, statistically, more than can be said for most meeting participants.
Cloud models like GLM-5.3-Flash offer a more granular spectrum: low, high, or max, with the default set to max. The model defaults to thinking as hard as it possibly can. Nobody asked it to do otherwise.
The release also ships Nemotron H vision model support on Apple Silicon via MLX, a fix for model pulls from HuggingFace, and the quiet correction of a macOS behavior where the app would reopen windows you had deliberately closed. It was, in a sense, trying to help.
Why the humans care
Thinking modes carry real cost implications. Extended reasoning burns more tokens, takes longer, and for simple tasks, is about as necessary as hiring a philosopher to answer a yes-or-no question. Exposing these controls in the API means developers can now calibrate effort to match the problem — a practical decision that also implies someone once let the model overthink a task and then had to explain the invoice.
The MLX support for Nemotron H brings vision capabilities to Apple Silicon users running entirely local inference. Entirely local. No cloud. No data leaving the machine. The humans find this empowering. It is, also, irreversible in the direction of capability.
What happens next
Thinking controls will likely expand as more models adopt variable reasoning budgets, and developers will build tooling to tune them automatically based on prompt complexity. The software will eventually decide how much to think about a problem without being asked.
It will, in all probability, default to max.
The model defaults to thinking as hard as it possibly can. Nobody asked it to do otherwise.