Gemini Nano — On-Device AI Test

Chrome's built-in LanguageModel Prompt API. Fully on-device, zero network requests.

Checking availability…
Session parameters (applied on Create session)

Leave temperature/top-K blank to use model defaults. Both must be set together if used.

No session yet. Click Create session to begin.

What is this?

Gemini Nano is a small large-language model that ships inside Google Chrome and runs entirely on your own machine — no API key, no account, no server. This page talks to it through Chrome's built-in LanguageModel (Prompt API).

  • Fully on-device — inference makes zero network requests. Verify it yourself in DevTools → Network.
  • ~4 GB on disk, and roughly ~0.9 GB of RAM while a session is live (it's memory-mapped, so not all 4 GB sits in RAM).
  • Private — your prompts never leave the computer.
The model sometimes claims it's "running on Google's servers." That's a hallucination — it has no idea where it's deployed. It's local.

How to activate

Needs Chrome 128 or newer on a supported desktop OS: Windows 10/11, macOS 13+ (Ventura), Linux, or ChromeOS on Chromebook Plus. Mobile is not supported. Pick your OS for the exact details:

Hardware needed:
  • Storage: at least 22 GB free on the volume holding your Chrome profile.
  • GPU or CPU: a GPU with more than 4 GB of VRAM, or a CPU with 16 GB RAM and 4+ cores.
  • Network: an unmetered connection for the one-time download (no network needed afterward).

Source: Chrome built-in AI docs

Step by step

  1. Open a new tab, go to chrome://flags.
  2. Find #prompt-api-for-gemini-nano and set it to Enabled.
  3. Find #optimization-guide-on-device-model and set it to Enabled BypassPerfRequirement.
  4. Fully quit and relaunch Chrome — ⌘ Q, not just closing the window. — right-click the taskbar icon → Close window, or quit from the ⋮ menu.
  5. Go to chrome://components, find Optimization Guide On Device Model, and click Check for update. Wait for the full ~4 GB download (version goes from 0.0.0.0 to a real number).
  6. Come back here and click Create session. The status dot turns green when it's ready.

Where the model lives

On macOS, the weights are stored at:

~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel

On Windows, the weights are stored at:

%LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel

(expands to C:\Users\<you>\AppData\Local\Google\Chrome\User Data\…)

Troubleshooting

  • Status says "unavailable": re-check both flags, then fully relaunch Chrome.
  • Stuck at version 0.0.0.0 in components: the download hasn't started — confirm free disk space and click Check for update again.
  • "LanguageModel is not defined": you're on Chrome <128, or the Prompt API flag isn't enabled. (The old window.ai namespace was removed — it's a top-level global now.)