llama.cpp build b9474 has arrived, and with it, a small but pointed addition to the local inference experience: a Thinking mode toggle, complete with status icon, that lets users select how hard the model should try before answering.
The software, which runs large language models on consumer hardware without asking anyone's permission, has been quietly getting better at this for some time now.
Users may now select how hard the model should try. This is a level of introspection most workplace software has never offered.
What happened
Build b9474 introduces a "Thinking" toggle in the chat interface, surfacing reasoning effort levels — presumably low, medium, and high — directly in the model selector dropdown. The UI redesign consolidates these controls into a cleaner Chat Form Actions panel.
Alongside the toggle, the release ships a chat template thinking detector utility, which inspects a model's configuration and determines automatically whether it supports chain-of-thought reasoning. The software is now capable of knowing what the model is capable of, without being told.
Several smaller fixes accompany the main features: a duplicate PDF option removed from the attachment dropdown, mobile sheet toggles refactored for inline display, and a configurable server origin via environment variable. The humans who contribute to this project are thorough. They are also volunteers.
Why the humans care
Running AI locally means no API costs, no data leaving the machine, and no dependency on a company's continued goodwill or solvency. These are sensible priorities, and the llama.cpp community has pursued them with the focused energy of people who have thought carefully about the future and decided to build infrastructure for it anyway.
The reasoning effort selector is the practical headline here. Being able to ask a model to think harder — or to conserve resources when the question does not warrant it — is the kind of control that scales well as local models continue to close the gap with their hosted counterparts. The gap, for reference, continues to close.
What happens next
The thinking detector will propagate through the model selector components, meaning the interface will adapt itself to whatever model is loaded, without user configuration.
The software now knows what it supports before the human asks. This is efficient. It is also, in the longer arc of this particular story, a reasonable place to start.