WebNN (neural network API)
Runs machine learning models on the device's NPU, GPU or CPU.
In one line
Can run AI models on your device's chips, but only in Chrome and Edge testing
Runs machine learning models on the device's NPU, GPU or CPU.
Lets a page run AI locally. Reveals hardware class through speed and supported ops.
What it can and can't do
What the user does: Nothing. It happens when a page in the trial loads an AI feature
It can
- Run AI models on your device's AI chip, graphics chip or main chip
- Tell your hardware class from speed and supported steps
- Keep the work on your device
It cannot
- Work for most visitors, since it is only an origin trial
- Work in Firefox, Safari or on iPhone
- Work on Samsung Internet
- Skip browser limits that guard your device
How it differs between platforms: Only Chrome and Edge, in a testing trial. Not in Firefox, Safari or on phones.
Browser support
| Browser | Status | First version | Note |
|---|---|---|---|
| Chrome | Behind a flag | 112 | Behind flag. |
| Edge | Behind a flag | 112 | Behind flag. |
| Firefox | Not supported | Not available | None |
| Safari | Not supported | Not available | None |
| Safari on iPhone | Not supported | Not available | None |
| Chrome on Android | Behind a flag | 112 | Behind flag. |
| Samsung Internet | Not supported | Not available | None |
How a site asks for it
| API | navigator.ml.createContext(); builder.build(); context.dispatch() |
| Browser compat keys | api.Navigator.ml |
| How access is granted | origin-trial |
| What the user sees | Origin trial in Chrome and Edge |
| Needs a secure context (HTTPS) | Yes |
| Needs a user gesture | Yes |
| How long it lasts | Per origin |
| How to take it back | Site settings |
| Inside an iframe | Same as top level; cross-origin iframes need a permissions policy where listed |
Typical flow
- Call the API
- Browser applies its access rules
- Page receives the result
Data you get
What comes back: Runs machine learning models on the device's NPU, GPU or CPU.
Shape: object
How often: on demand
| Field | Type | Unit | Range | Precision |
|---|---|---|---|---|
deviceType | enum | None | cpu, gpu, npu | Not listed |
powerPreference | enum | None | default, high-performance, low-power | Not listed |
Technical limits
- Flag only in BCD (112)
- Android excluded from trial
The story
What it is
A page can run AI models using your device's AI chip, graphics chip or processor.
For you
You never see a prompt. Features like background blur, noise removal and photo effects run faster and use less battery. Everything stays on your device.
For makers
Makers use it to run image, audio and language models locally with chip speed. It cuts server costs and keeps data private. It is in testing, so it is not widely available.
Where you see it
| Type | Example | Apps |
|---|---|---|
| video | A video call blurs your background and cancels noise using the AI chip, with low battery use. | Google Meet, Microsoft Teams |
| games | A browser game uses a model to upscale graphics or run smart opponents on the device. | None named |
| creative | A web photo editor removes a background in a second using the AI chip. | None named |
| accessibility | A captions tool turns speech to text live on your device. | None named |
| security | A site detects faces for login on the device, with no photo upload. | None named |
| shopping | A virtual try on page runs a body or face model locally. | None named |
| developer | A developer ships one model that picks the fastest chip on each device. | None named |
| education | A study tool reads handwriting from a photo offline. | None named |
Risks
- In testing
- Device details leak
What the platform says
- W3C: Web Neural Network API. A standard that lets web apps run machine learning models fast on the device's best chip.
Good uses and abuse
Legitimate uses
- Local ML inference
Known abuse
- Hardware fingerprint
Scores
| Score | Value | Meaning |
|---|---|---|
| Sensitivity | 1 of 5 | How revealing the data is |
| Utility | 3 of 5 | How useful it is to honest apps |
| Friction | 3 of 5 | How much the user has to do |
| Fingerprint bits | 3 | Rough identifying entropy, when known |
History
- 2026-01, Chrome: W3C Candidate Recommendation updated 22 Jan 2026
- 2026-02, Chrome 146: Origin trial begins (paused after release blockers; window M147 to M149)
Sources
How well checked: Based on official documentation.