Hi there,
I’d like to discuss direction rather than any specific implementation: is there any plans to make local AI models first class citizens in the GNOME (and freedesktop in general) Desktop?
Some examples where this would be useful. Matching faces in a photo library to address book entries (Evolution Data Server / Contacts), so contacts gain a portrait automatically. A photobook application using a local LLM and image models to categorise photos by content or occasion and draft captions. Dictation wanting speech-to-text, a notes app wanting summarisation, a viewer wanting on-device OCR. Different capabilities, one underlying need — and for personal data like contacts and photos, on-device is exactly where it belongs.
Today each app carries its own stack: digiKam has its face engine, LLM apps pull in Ollama, dictation tools bundle Whisper. So users might need to re-download the same multi-gigabyte models per app, and sandboxed (Flatpak) apps must request broad GPU access that weakens the isolation we value.
The proprietary platforms have converged on a system-service model: Android’s AICore manages models centrally, Apple ships an OS-provided model behind a stable API, Windows exposes capability-oriented AI APIs over a hardware-abstraction layer. In each case the model lifecycle lives in one place, and apps request a capability rather than a named model.
So is there any plans/initiatives to have basic building blocks for using AI models on a desktop?
Best,
Tom