Edge opens Aion local AI to live-site trials, with limits to check

Microsoft has opened a live-site trial of the Prompt API in Edge and made Aion Instruct the browser’s built-in small language model. Its November 19 announcement separates that trial from a second, more experimental interface for classifying text. For website owners, the practical change is a route to test local AI with real visitors, subject to device and trial requirements.
The >Edge team’s announcement (opens in a new tab) says Aion replaces Phi-4-mini. Microsoft also released the model’s weights on Hugging Face. Its model card lists about 2.3 billion parameters and an MIT license. Those facts describe the available model, not a guarantee that a particular laptop can run a website’s feature successfully.
Three different levels of availability
The model release, the Prompt API origin trial and the Decisions API preview are separate steps. An origin trial lets registered sites test an experimental browser capability with their users. It does not make the interface a settled standard across browsers.
Microsoft’s developer documentation describes the Decisions API as a preview for Edge Canary. Rather than composing a response, it scores predefined answers to questions about text and returns probabilities and a confidence value. A site could use that structure to route feedback or sort requests. A confident answer can still be wrong; the receiving application must decide what evidence is sufficient before acting.
Local processing needs a separate cloud decision
The Prompt API documentation says inference runs on the user’s device. After the initial model download, that local operation can work offline. The same documentation lists hardware and storage conditions, so availability needs to be checked rather than assumed from the browser name.
A website can also be designed to use a cloud provider when local processing is unsuitable. That is a separate data path. Describing the local model as private does not establish what the rest of a website collects or sends elsewhere. A useful implementation should explain the fallback before a visitor entrusts it with sensitive text.
Test the language as well as the device
Microsoft’s model card lists English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese. Hebrew, Arabic and Russian are absent from that list. The release therefore does not establish supported performance in those languages.
The card also warns about factual mistakes, reasoning limits and adversarial inputs. For a site testing the new capability, representative examples in the intended language matter more than a general speed claim. Start with a bounded task, keep a usable path when local AI is unavailable, and check the output before turning a classification into a consequential action.


