Chrome ships a growing set of on-device AI APIs — Prompt, Summarizer, Translator, Language Detector, Writer and Rewriter — that run a local model instead of calling a server. Most of them are still gated behind an origin trial or a Chrome flag, so knowing what each availability state actually means, and what happens before you trigger a multi-gigabyte download, saves you from guessing.
Why availability differs from browser to browser
These APIs ship progressively: some are stable in recent Chrome desktop builds, others require enabling a flag at chrome://flags or joining an origin trial for a specific site, and none of them currently work in Chrome on mobile.
Even on a qualifying Chrome version, the underlying model (Gemini Nano) has its own OS and hardware requirements — enough free disk space and a supported platform (Windows, macOS or Linux) — so the same Chrome version can report different availability on different machines.
The states an availability check can return
Calling an API's availability() method resolves to one of a few states: unavailable means the browser supports the API but your device doesn't currently qualify for it; downloadable means everything checks out but the model itself hasn't been pulled to your device yet; downloading means a pull is already in progress; and available means the model is ready to use right now.
A missing API entirely — no global for it in the browser at all — is a separate case from all of those: it means this Chrome build doesn't expose that API, whether because it's too old, the flag isn't enabled, or the API hasn't shipped there yet.
What actually happens when you download a model
A "downloadable" status doesn't mean anything has downloaded yet — that only happens when you call the API's create() method, which is also how you get the session or instance you actually use. Depending on the API and model, that download can range from a couple hundred MB to a few GB, and only needs to happen once per device since Chrome caches it for every site that uses the API afterward.
Because that's a real, potentially large download, a page built around these APIs should make it an explicit step a visitor opts into — never something that fires automatically just because a page loaded or a tab was opened.
Trying it yourself
The Browser AI Tester on this site runs every one of these checks live in your browser and lets you try each API — including its own explicit "Download model" step, separate from running it — without sending anything you type to a server.