Picture this: you open a web page, highlight a long article, and ask your browser to sum it up. Nothing leaves your computer. No spinner, no server round trip, no data sent to a distant data centre. The answer appears in a blink, written by a model that lives inside the browser itself.
That is the promise Microsoft made on November 19, 2026, when it announced that Microsoft Aion Instruct is now generally available as the built-in small language model of Microsoft Edge. The 2.3-billion-parameter model replaces Phi-4-mini, the model Edge has shipped with until now, and it is also published as open weights on Hugging Face for anyone to download.
What Is Microsoft Aion Instruct?
Microsoft Aion Instruct is a small language model built by the Windows Applied Sciences Group. It is an instruction-tuned model with about 2.29 billion parameters and a 32,768-token context window, which is enough to hold a long document, a chat history, or a chunk of code at once.
The weights are published on Hugging Face under the MIT licence, as microsoft/Aion-1.0-Instruct-2.3B. That matters because anyone can download the model, inspect it, and run it outside Microsoft’s products. It is a genuinely open release, not a demo locked behind an API key.
Microsoft describes Aion Instruct as a model for everyday text work: writing assistance, summarisation, instruction following, and structured output. It is not trying to be a frontier giant like Google’s Gemini 4 Carbon. It is trying to be the model that is already there when you need it, running quietly on your own machine.
Why Aion Replaces Phi-4-mini
Edge has had a built-in model for a while. Phi-4-mini powered the browser’s on-device features, and before that, other small Microsoft models did the same job. Aion Instruct is the next step in that line, and Microsoft says it is smaller and faster than Phi-4-mini while delivering comparable quality.
The swap is not just about the browser. The same model is also replacing Phi Silica as the on-device language model of Windows itself. Microsoft plans to roll it out to Windows Insider devices in December 2026, with retail Copilot+ PCs following in January 2027.
That gives Aion Instruct a much bigger stage than a browser feature. One model, shared by Edge and Windows, means developers can build on-device features once and reach both.
The Prompt API Origin Trial: Websites Can Now Call the Model
The bigger news for developers is the Prompt API. Until now, the API was a developer preview limited to Edge Canary and Dev, switched on through a flag. With the new origin trial, a website can register, add a token to its pages, and call the built-in model from JavaScript on a live site for real users.
The idea is simple and radical at the same time. A page checks whether the model is available, then asks it to write, summarise, or reformat text. The work happens on the visitor’s device. The text never has to travel to a server, which is a genuine privacy win for anything sensitive.
Microsoft wants feedback before the API goes further. The trial should surface real-world problems: how the API feels to use, how the model performs on ordinary hardware, and whether the same interface could work across browsers. That feedback will feed into W3C Web Machine Learning discussions, where the Prompt API is being considered for standardisation.
Chrome already offers something similar: its Prompt API, backed by Gemini Nano, is available to ordinary websites. Edge is now catching up, and the two browsers are converging on the same idea — the browser as a platform for local AI, not just a window onto cloud AI.
A Decisions API Joins the Party
Microsoft also announced an experimental Decisions API, powered by the same Aion Instruct model. It answers yes-or-no, multiple-choice, and scoring questions about a piece of text, without generating any prose at all.
Think of it as a fast judge rather than a writer. A site could ask whether a comment is spam, score a product review, or sort incoming messages, all on the device. The API is only a developer preview in Edge Canary and Dev for now, so it is the least mature of the three announcements.
The timing is interesting. Google announced its own Gemini Decisions API days earlier, and OpenAI shipped a Decisions API in October. Structured, scored answers are becoming a pattern across the industry, because many real jobs do not need a paragraph. They need a verdict.
What On-Device AI Means for Your Privacy
The privacy argument is Microsoft’s strongest one. When a model runs on your device, your text stays on your device. A summariser that never phones home is a different product from one that uploads your document to a cloud server first.
That matters most for the boring, sensitive stuff of daily life: work documents, medical forms, private messages. A browser that can summarise those locally, without a round trip, is genuinely useful in a way that a cloud model cannot match without a trust leap.
There is a catch, of course. On-device AI only works if the device can run it, and a model living on your machine still needs guardrails. Microsoft’s own researchers have shown how a handful of altered neurons can turn an AI agent rogue, which is one more reason on-device models need provenance checks and clear defences. Microsoft’s documentation lists Windows 10 or 11, macOS 13.3 or later, and qualifying 64-bit Linux systems with Vulkan GPU drivers. The profile drive needs at least 20 GB free, and the GPU needs at least 3 GB of video memory.
Those requirements are not extreme for a modern laptop, but they do exclude older machines and thin clients. Aion Instruct is also designed to run on less capable GPUs and even CPU-only machines, which is part of why Microsoft picked a 2.3-billion-parameter model instead of something bigger.
The Catch: Microsoft Has Not Published the Fine Print
For all the detail in the announcement, some basics are still missing. Microsoft has not published exact hardware thresholds, the download size of the model, or the precise Edge versions that will receive it. The announcement also gives no absolute speed figures for common tasks.
That is worth noting because the whole pitch is about speed and convenience. A model that is “smaller and faster” is only a win if it is fast enough on the machines people actually own.
The origin trial itself is temporary by design. Tokens expire after six weeks by default and can be renewed, and the feature only reaches users on registered origins. Anyone building on it should treat it as an experiment, not a foundation, until Microsoft commits to a stable release.
What Happens Next
The next milestone is December, when Aion Instruct starts reaching Windows Insider devices as the new on-device model for Windows. January brings the wider rollout to Copilot+ PCs, and with it, a much larger audience for on-device features built on the same model.
Developers can register for the Prompt API origin trial now, read the documentation, and try the playground sample Microsoft published alongside the announcement. The company is collecting feedback on GitHub, and that feedback will shape whether the Prompt API becomes a web standard or stays a Microsoft experiment.
The bigger picture is that the browser is becoming an AI platform in its own right. First came cloud assistants in a sidebar. Now the model is moving into the browser itself, small enough to live on your machine and open enough for any website to use. Microsoft’s bet is that the next wave of web AI will not be a chatbot you visit. It will be a capability your browser already has.
Microsoft announced the changes on November 19, 2026, on the Microsoft Edge Blog. The model is available now on Hugging Face, and the origin trial is open for registration.



