Google Chrome Built-in Multimodal Model (Gemini Nano) Local Setup Guide
Google Chrome Built-in Multimodal Model (Gemini Nano) Local Setup Guide
This guide walks you through how to download, configure, and call Google Chrome’s built‑in lightweight multimodal model Gemini Nano in a local environment, enabling offline inference and image understanding.
1. Environment Setup and Experimental Flags
1.1 Check model component status
First, open chrome://components and search for the keyword “model” to verify whether the model component is enabled:
1.2 Enable the Prompt API
Navigate to chrome://flags and enable the flags related to the on‑device model and Prompt API:
1.3 Configure multimodal access
Enable access to the Gemini Nano model and switch it to multimodal mode to support image parsing:
Tip: GLIC is Google’s internal codename for Google LLM In Chrome.
After configuration is complete, click Relaunch to restart the browser.
1.4 Configure on-device internal services
Open chrome://on-device-internals/. On first visit, you’ll see an access control prompt:
Click Connect and switch the status to Enable. The page will then redirect to the management dashboard:
2. Model Download and Deployment
2.1 Check model status
Click the Model Status tab to confirm that your device hardware and memory meet the requirements. Initially, you won’t see any model downloads in progress:
2.2 Force update and download models
Click Reset to force an update, then download the small helper models (used for image understanding, prompts, etc.):
At this point, a model event entry will appear. Click Check and Chrome will call the backend service to download the on‑device Gemini Nano v3 model:
2.3 Verify download completion
When the download finishes, you’ll see a success message. The local small model is now ready for offline use:
You can now use the model for simple tasks like webpage error checking. In the example below, the model still works in offline mode and can highlight issues:
3. Multimodal Capability Test
3.1 Load the multimodal UI
Go back to the main page. After all models finish downloading, click Load Default. The interface will now support uploading images and videos:
3.2 Image understanding test
Upload a test image and let the model recognize it. For example, you can upload a photo of “a person holding a camera and shooting in the sunset”:
From the test results, Gemini Nano can correctly understand simple visual scenes.
Note: Audio recognition wasn’t tested here due to format compatibility issues. If you’re interested, you can experiment with the speech recognition capabilities yourself.
4. Where the Model Files Are Stored
All downloaded model files are stored in the following path:
1 | %LocalAppData%\Google\Chrome SxS\User Data |
Note: In this example, we’re using Google Chrome Dev Channel, so the path contains
Chrome SxS. The path may differ for Stable or Beta builds. You can also manually copy model files into this directory if needed.
Summary
By following the steps above, you’ve successfully deployed Google Chrome’s built‑in Gemini Nano multimodal model locally, and can use it for offline image understanding, text analysis, and other AI tasks. This client‑side AI approach has several advantages:
- Privacy: Data is processed locally without being sent to the server.
- Low latency: No network round‑trips, resulting in faster responses.
- Offline availability: AI functions remain usable even when the network is disconnected.
- Cost optimization: Reduces server‑side inference load and associated costs.
