Microsoft is swapping out the AI model in its browser. With Aion Instruct, Microsoft Edge gets a new built-in language model that replaces Phi-4-mini. The 2.3-billion-parameter model has also been published as an open-weight model under the MIT licence on Hugging Face. For web developers, the second announcement is at least as important: the Prompt API, which lets websites call the local model from JavaScript, is entering an origin trial and can now run on live sites for the first time. An experimental Decisions API comes on top.
Aion Instruct replaces Phi-4-mini in Edge
Microsoft presented the changes on November 19 on the Edge developer blog. According to the Edge team, Aion Instruct is smaller and faster than Phi and runs on many more devices, while delivering comparable quality for common web tasks. Those include text synthesis and generation, summarisation, instruction following and structured outputs. For comparison, Phi-4-mini, which Edge has used since the Prompt API launched, has around four billion parameters.
The switch comes later than planned. In June, Microsoft showed a preview of Aion in the Edge Canary and Dev channels and said the open-source release on Hugging Face would follow in July. Now the model is officially available. Edge only downloads it when a site or feature first needs it, and only on devices that can run it well. Users can turn off on-device AI in settings, which removes the model and stops future downloads.
Windows is set to get Aion as well. According to the model card on Hugging Face, a further optimised version, paired with a Microsoft content moderation model, will become the operating system’s language model by the end of January 2027. The Hugging Face release is meant as a preview and a step towards transparency. The unquantised model comes first, quantised versions are to follow.
What the model card reveals
Aion Instruct comes from the Applied Sciences Group in Microsoft’s Windows division. According to the model card, its training data extends the data used for Phi Silica, the previous Windows language model. Multilingual data makes up only about 6 percent; eight languages are officially supported, including German.
- Parameters: about 2.29 billion, 28 layers
- Context: 32,768 tokens
- Attention: 21 sliding-window layers (window 2,048) plus 7 full-attention layers
- Knowledge cutoff: December 2025
- Licence: MIT; formats PyTorch, ONNX and GGUF
Microsoft ran the benchmarks itself; there are no independent measurements yet. Aion scores 55.15 percent on MMLU, 33.35 percent on MMLU-Pro, 74.91 percent on GSM8K and 83.90 percent on instruction following (IFEval). Those are small-model numbers, and Microsoft says so openly: Aion was not trained for world knowledge, can miss details, make misstatements and fail at complex multi-step reasoning. It may also be susceptible to adversarial attacks. Microsoft explicitly names cross-prompt injection (XPIA) as one of the metrics to improve in future releases.
Prompt API origin trial: AI in the browser without the cloud
Until now, the Prompt API was only available as a developer preview in Edge Canary and Dev. With the origin trial, site owners can use it on their live websites. They register their domain, receive a token and serve it in a meta tag or an Origin-Trial HTTP header. Edge then enables the feature for visitors automatically. According to Microsoft’s documentation, these tokens expire after six weeks by default and need to be renewed.
Using it in code is straightforward. The developer documentation describes the LanguageModel object: LanguageModel.availability() tells a page whether the model is ready, needs to be downloaded or is unavailable. LanguageModel.create() opens a session, and session.prompt() or session.promptStreaming() return the response. The monitor option tracks the initial download.
A minimal example based on the documentation looks like this:
const availability = await LanguageModel.availability();
if (availability !== "unavailable") {
const session = await LanguageModel.create();
const result = await session.prompt("Summarise this text in two sentences: …");
console.log(result);
}
One point matters for companies: the documentation names the GenAILocalFoundationalModelSettings and ComponentUpdatesEnabled policies, which administrators can use to block the model download. Anyone running Edge in managed environments should check whether on-device AI is wanted there at all before business units build web applications on top of it.
The hardware requirements are not trivial. According to the documentation, the preview currently needs Windows 10 or 11, macOS 13.3 or later, or 64-bit Linux with Vulkan drivers and Edge 159, at least 20 gigabytes of free space on the profile drive, a GPU with at least 3 gigabytes of VRAM and an unmetered connection. If free storage drops below 5 gigabytes, Edge deletes the model again. Websites therefore always need a fallback.
Decisions API: classification without text generation
The Decisions API is new and far more experimental, available for now only in Edge Canary and Dev. It answers yes/no, multiple-choice or scoring questions about a piece of text and returns probabilities and a confidence value, computed in a single pass without generating text. Microsoft cites detecting user intent, sorting and routing user-generated content and scoring items in real time as examples. In the Edge team’s sample, the API analyses a product review for topic, satisfaction and whether the customer needs help, all at once. Both the API and its implementation are expected to change.
Context: hybrid AI becomes a web platform topic
Microsoft calls the approach “hybrid intelligence”: sites should use the on-device model where it is available and well-suited, and fall back to a cloud provider otherwise. For web developers that has tangible benefits. Local inference costs nothing per call, avoids server round trips and keeps text on the device. The drawbacks are just as clear: a 2.3-billion-parameter model is no substitute for large cloud models such as Microsoft’s own MAI Code 2 family, and the Prompt API is not a web standard yet.
That is exactly what the origin trial is meant to address. Feedback on API design, model performance and interoperability is to feed into discussions with other browser vendors, since the Prompt API is being considered for standardisation. Until then, anyone adopting it should use feature detection and plan a fallback. How quickly browser platforms change is also shown by the removal of XSLT in Chrome 158. Small models such as Aion or Meta’s Muse Ember show where things are heading: routine tasks move onto the device, the heavy cases stay in the cloud.




