On November 19, 2026, Microsoft released Aion Instruct, the new built-in language model of the Edge browser. It has 2.3 billion parameters, replaces Phi-4-mini, and runs on the visitor's device instead of in the cloud.
For website operators, the real point lies in the Prompt API. Since November 19, it can be used on a public website as an Origin Trial, that is, as a time-limited test phase. Until then, it was only available in the preview versions Edge Canary and Dev.
What can a website do with the model?
A page calls the model via JavaScript, for example for text drafts, for analyzing inputs, or for responses in a predefined structure. According to the documentation, the inputs do not leave the device. The operator incurs no fee for a cloud service. Microsoft itself warns that a model of this size can respond incorrectly or unreliably.
As a developer preview in Edge Canary and Dev, there is also the Decisions API. With it, a page poses yes/no, multiple-choice, or rating questions about a text to the model and receives probabilities along with an indication of how certain the model is, without the model generating text. Microsoft cites sorting customer reviews or recognizing what a user intends to do as examples. The API is experimental and may change.
On which devices does the model run?
According to the Edge documentation, the model requires Windows 10 or 11, macOS 13.3 or later, or 64-bit Linux with a Vulkan driver, and on Linux additionally Edge version 159 or later. In addition, it requires at least 20 GB of free space in the Edge profile, 3 GB of video memory, and a connection without a data limit. On a metered connection, Edge does not download the model.
The model needs these requirements
Information from the Edge documentation on the Prompt API, as of November 22, 2026
- Windows 10 or 11, macOS 13.3 or later, Linux with Vulkan and Edge 159 or later
- At least 20 GB of free space in the Edge profile
- Video memory of 3 GB or more
- A connection without a data limit for the first download
- Local AI turned on in the Edge settings
- Blocked if a company policy blocks the download
Experimental feature: Microsoft states the values for the developer preview
Sources: Microsoft Edge Blog, Microsoft Learn
If there is less than 5 GB of free space, Edge deletes the model again. The browser only starts the download when a page or feature needs the model for the first time. This only happens on devices where it runs well, as Microsoft writes in the Edge blog. Users turn off local AI in the Edge settings under "Copilot and AI," after which Edge removes the model and does not download it again. A company policy can also block the download.
The blog describes Aion as smaller, faster, and able to run on more devices than Phi-4-mini. Whether the minimum values in the documentation change with Aion, and for which countries and languages the model is intended, Microsoft does not say.
What this means for you
To use the Prompt API on your website, you check in the code with LanguageModel.availability() whether the model is available. Visitors without the model, for example with weak hardware or local AI turned off, still need a fallback solution that works without the API.
To try it out, from Edge 158 onward you turn on the "Prompt API" flag under edge://flags. Under edge://on-device-internals you can see the device class. From the device class "Medium" upward, the model should run on your computer.
In your own Edge browser, you turn off local AI in the settings so that the model does not occupy 20 GB. For sensitive tasks, such as price information or legal texts, you are better off not relying on the answers of a model with 2.3 billion parameters.
Microsoft presented the new features in the Edge blog on November 19, 2026.
Sources