Your Browser Now Has an AI Model Built In — and Websites Can Use It
Microsoft just turned Edge into something new: a browser with a small language model living inside it, which websites can now call directly. Announced November 19, 2026, on the Microsoft Edge Blog, the change has two par
Microsoft just turned Edge into something new: a browser with a small language model living inside it, which websites can now call directly. Announced November 19, 2026, on the Microsoft Edge Blog, the change has two parts: Aion Instruct is now Edge's built-in small language model, replacing Phi-4-mini, and the Prompt API has moved from a developer preview in Canary and Dev builds into an origin trial, so registered sites can use it on live websites.
A small model with a big job
Aion Instruct has 2.3 billion parameters, and Microsoft says it is smaller and faster than the model it replaces while keeping comparable quality. It is built for a short list of jobs: text synthesis and generation, summarisation, following instructions, and structured outputs. Those sound modest, but they cover most of what a website would ask a small model to do — draft a reply, shrink a long page into a summary, or turn loose text into tidy structured data.
The model is also out in the open on Hugging Face as microsoft/Aion-1.0-Instruct-2.3B, under the MIT licence, with a 32,768-token context window — so developers can test it outside the browser and compare it with other small models.
The Prompt API opens up
Until now, the Prompt API was limited to preview builds, which meant only testers could try it. An origin trial changes that: sites that register can call the on-device model from JavaScript on real, live pages with real visitors.
One detail makes the whole thing practical. The model downloads once, the first time it is used, and then it is shared across sites. A visitor who has already used one site that calls the Prompt API will not download it again for the next one. And because the model runs on the device, features like writing help and page summaries work without sending text to a server.
A second API for simple answers
Microsoft also showed off an experimental Decisions API, available in Canary and Dev builds only. It answers yes or no questions, picks from multiple-choice options, or returns a score with probabilities — and it generates no text at all.
That sounds limited, and it is meant to be. Plenty of web features only need a decision — is this spam, which category fits, how likely is this a match? — and a light API that returns a choice and a probability handles those jobs without the cost of full text generation.
The hardware bar
On-device AI still needs decent hardware. The Prompt API asks for Windows 10 or 11, macOS 13.3 or later, or 64-bit Linux with Vulkan GPU drivers, plus Edge 159 or newer. On top of that, you need 20GB of free storage, 3GB or more of VRAM, and an unmetered connection.
Watch the storage rule: if free space drops below 5GB, the model gets deleted.
Windows gets the same model
The story does not stop at the browser: the same model will replace Phi Silica as the on-device model in Windows, starting with Windows Insider devices in December 2026 and reaching retail Copilot+ PCs in January 2027.
There are limits, and Microsoft's own model card states them plainly: a 2.3B-parameter model can miss details and fail at complex multi-step reasoning. That is the honest trade of small models — quick and cheap to run, good at focused jobs, and not a stand-in for a large model on hard problems.
The full announcement is on the Microsoft Edge Blog, and the API is documented in the Prompt API docs. If this origin trial goes well, the browser starts to look less like a window onto AI services and more like a place where the AI already lives.
Read the full article with all the details on Rohan Tech Official: https://rohantechofficial.com/tech-news/microsoft-aion-instruct-edge-prompt-api/
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.