Microsoft Puts Aion Instruct Inside Edge — A 2.3B AI Model Websites Can Call On-Device

Microsoft has put a new small language model inside Edge, and this time websites can call it directly. On November 19, 2026, the company announced Aion Instruct on the Microsoft Edge Blog: a 2.3B-parameter model that now ships as the browser’s built-in small language model, taking over from Phi-4-mini. In the same announcement, Microsoft moved the browser’s Prompt API out of its Canary and Dev developer preview and into an origin trial, so registered sites can start calling the on-device model from JavaScript on live websites.
That is a real shift in how browser AI works. Until now, a site that wanted AI features had to send text to a server, or ship its own model and hope the visitor’s machine could handle it. With Aion Instruct built into Edge, the model lives in the browser itself, and any registered site in the origin trial can ask it to write, summarise, or structure text without a round trip to a cloud service.
What Aion Instruct Is
Aion Instruct is a small language model tuned for everyday text work. Microsoft says it is smaller and faster than Phi-4-mini while holding comparable quality across the jobs it targets: text synthesis and generation, summarisation, instruction following, and structured outputs. Those four jobs cover a lot of what websites actually ask a small model to do, from drafting a reply to turning a wall of text into a short summary or a neat JSON object.
The model is also available as open weights on Hugging Face under the name microsoft/Aion-1.0-Instruct-2.3B, released under the MIT licence, with a 32,768-token context window. That open release matters for developers who want to test the model outside Edge, or compare it with other small models on their own terms. Inside Edge, though, it plays a different role: it is the default small model the browser itself offers to web pages.
Replacing Phi-4-mini is a statement of intent as much as a swap. Phi-4-mini was Microsoft’s earlier small model for the browser, and Aion Instruct takes its place as the built-in option. For users, the change should be invisible most of the time. For developers, it means the model they can call from a web page has changed, and the one now in place is built to run faster on ordinary hardware.
The Prompt API Moves to an Origin Trial
The Prompt API is the JavaScript interface that lets a web page talk to the on-device model. It has been in developer preview in Edge Canary and Dev builds, where only testers and early builders could try it. Moving to an origin trial opens it up: registered sites can now call the model on live sites, with real visitors, instead of only in test builds.
The model downloads once, on first use, and is then shared across sites. That detail matters. A visitor who has already used one site that calls the Prompt API will not need to download the model again for the next site. The download happens once, the model stays on the device, and each site that has registered for the trial can use it.
Because the model runs on the device, a site can offer AI features that keep working with the text right there in the browser. Writing help, summaries of long pages, and structured answers can all run locally, which also means the feature does not depend on a server being up or a network call coming back in time.
A Second, Smaller API: Decisions
Alongside the Prompt API, Microsoft introduced an experimental Decisions API, available in Canary and Dev builds only. It does one narrow thing: it returns a yes or no answer, picks from a set of multiple-choice options, or gives a scoring answer with probabilities. It does no text generation at all.
That narrow scope is the point. Plenty of web features do not need a paragraph of generated text. They need a call: is this message spam, which category does this item belong to, how likely is this a match? A small API that answers with a choice and a probability fits those jobs without the cost of running full text generation, and it gives developers a lighter tool for classification-style work inside the browser.
What Your PC Needs to Run It
On-device AI still needs real hardware, and Microsoft has spelled out the bar. The Prompt API needs Windows 10 or 11, macOS 13.3 or later, or 64-bit Linux with Vulkan GPU drivers, running Edge 159 or newer. On top of that, the machine needs 20GB of free storage, 3GB or more of VRAM, and an unmetered connection.
The storage rule has a sharp edge: if free space drops below 5GB, the model is deleted. That is worth knowing if you run a machine that is often close to full, since the model would then need to download again on first use. The unmetered connection requirement also tells its own story: the first download is not something Microsoft wants happening over a capped mobile connection.
The Same Model Is Coming to Windows
Aion Instruct is not only an Edge story. Microsoft says the same model will replace Phi Silica as the on-device model in Windows itself. The rollout starts with Windows Insider devices in December 2026, followed by retail Copilot+ PCs in January 2027.
That puts one model at the centre of Microsoft’s on-device AI on the PC: the same 2.3B model in the browser and in the operating system. For developers, that kind of shared base is useful. A feature built around the model in Edge and a feature built around it in Windows would be working with the same behaviour and the same limits.
The Honest Limits
Microsoft’s own model card keeps expectations in check. It warns that a 2.3B-parameter model can miss details and fail at complex multi-step reasoning. That is a fair warning, and it fits what small models are: quick, cheap to run, and good at focused jobs, but not a stand-in for a large model on hard problems.
It also sets the terms for how the Prompt API should be used. Summaries, short drafts, and structured outputs sit inside the model’s strengths. Long chains of reasoning, or answers where a missed detail changes the result, sit outside them. Sites that treat the on-device model as a helper for small, well-scoped tasks will get more from it than sites that ask it to do the heavy lifting.
Small on-device models have been a slow build across the industry, and Microsoft’s move here is one more step in that direction. If you want a sense of how local models are shaping up elsewhere, our look at Inkling 2 Mini, a local AI model built to run on-device, covers similar ground. The security side matters too: our report on the Microsoft Agent Framework PyPI compromise is a reminder that the supply chain behind AI tooling needs watching. For the full announcement, see the Microsoft Edge Blog post on Aion Instruct and the Prompt API, and the Prompt API documentation for how the API itself works.
What happens next is worth watching on two fronts: how sites use the origin trial once real visitors start hitting it, and how the Windows rollout lands in December and January. If the model holds up in daily use, the browser may start to feel less like a window onto AI services and more like a place where the AI already lives.
