Gemini Nano is the small AI model that Chrome downloads and runs on your own computer. Chrome uses it for some of its own features, such as scam warnings, and web developers can use it through Chrome’s built-in AI APIs, for example to summarize text without sending it to a server. You manage it in Chrome under Settings, AI innovations, On-device AI.
This page covers what the model needs, how to turn it off, and what developers can build with it. For Chrome’s other AI tools, see our overview of Chrome’s AI features.
What it does for you
Google’s help article Manage on-device Generative AI models in Chrome says on-device models “might be used” by features such as help with writing or rephrasing text, scam warnings and page summaries. Data is processed on your device. Some of Chrome’s AI features do not use the on-device model at all and keep working if you remove it.
Google’s AI in Chrome page says Gemini Nano lets Chrome detect scams that have not been seen before.
When Chrome downloads it
Chrome may download the model in the background on eligible devices, so the features that need it are ready. Google’s help lists these conditions:
- An unmetered network connection. Google’s developer docs note that Wi-Fi and Ethernet are usually unmetered, while mobile data is often metered.
- About 20 GB of free disk space.
- A device powerful enough to run it. Google says the exact requirements may change.
The setting is per user and per device, so people who share a computer each keep their own choice.
How to turn it off
- Click the three-dot menu, then Settings.
- Click AI innovations.
- Turn On-device AI off.
Chrome then deletes the model files, features that rely on them stop working, and Chrome will not download the model again while the setting is off. Google says this toggle is the only supported way to manage the model, and that deleting the files by hand is not recommended and may not stop future downloads.
On managed computers, administrators use the GenAILocalFoundationalModelSettings policy. Set to “Do not download model”, Chrome does not download the model and deletes any copy already downloaded.
For developers: the built-in AI APIs
With Chrome’s built-in AI, a website or extension can run AI tasks with models that the browser downloads and manages. Google’s developer docs list the APIs and their status on Built-in AI APIs:
- In stable Chrome: the Translator API, the Language Detector API, the Summarizer API and the Prompt API.
- In trials: the Writer and Rewriter APIs and the Proofreader API. Check the status table before you rely on them, because it changes.
The Translator and Language Detector APIs use expert models. The others use Gemini Nano.
Hardware your users need
Google’s Get started with built-in AI page sets these requirements for the APIs that use Gemini Nano:
- Operating system: Windows 10 or 11, macOS 13 or later, Linux, or ChromeOS on Chromebook Plus devices. Chrome for Android and iOS is not supported for these APIs.
- Storage: at least 22 GB free on the drive that holds the Chrome profile. If free space falls below 10 GB after the download, Chrome removes the model.
- GPU or CPU: a GPU with more than 4 GB of video memory, or 16 GB of RAM and at least 4 CPU cores.
- Network: an unlimited or unmetered connection, needed only for the first download.
Google’s developer page states: “No data is sent to Google or any third party when using the model.”
How the model download works
The first call to a create() function of an API that uses Gemini Nano, such as Summarizer.create(), starts the download. Call availability() first: it returns "unavailable", "downloadable", "downloading" or "available". If the model still has to be downloaded, the user must have interacted with the page, for example with a click or key press, before you can create a session.
Google’s page on model management explains the rest. Chrome picks a larger or smaller variant of the model based on how fast the device’s GPU is, or falls back to the CPU. Interrupted downloads resume. Updates are full new downloads that swap in without downtime. Chrome can delete the model at any time, even in the middle of a session, when disk space runs low, so your code should handle an API becoming unavailable.
Check the model and debug prompts
Open this page to check the model’s current size and read its event logs:
chrome://on-device-internals
Google’s guide to debugging Gemini Nano shows how to use the Event Logs tab and the Dump button. Administrators can turn the APIs off for websites with the BuiltInAIAPIsEnabled policy, in the same policy list as the download setting above.