The News, Explained
Microsoft announced on November 19, 2026 that Aion Instruct is replacing Edge’s built-in language model and that websites can test the Prompt API with visitors on live sites. An API is a set of rules through which site code calls a browser capability. Model general availability, site-specific trials and an early experimental feature were announced together, so their access conditions need to be separated. Source
Aion is the small language model replacing Phi-4-mini in Edge. It has about 2.3 billion parameters, the numerical values holding learned information within a model. Microsoft names text generation, summarization, following instructions and producing structured responses as web tasks for it, and provides open weights on Hugging Face. Its model card also describes limits: a small model can miss details, produce misstatements and struggle with complex reasoning. Source Source
To test the Prompt API with real site visitors, a developer registers the site’s origin for an origin trial and applies a trial token. The token marks the registered site as eligible for the experimental capability. The current trial page lists April 18, 2027 as its expiration date. The feature may change or be suspended, and permanent availability after the trial is not assured. Model availability and every website immediately offering the same AI feature are therefore distinct situations. Source Source
Current documentation lists Windows 10 or 11, macOS 13.3 or later, and 64-bit x64 Linux with Vulkan GPU drivers; Linux also requires Edge 159 or later. Hardware needs at least 3GB of VRAM, the GPU’s own memory, and at least 20GB free on the drive containing the Edge profile. That 20GB is a free-space requirement rather than a model-file size. Downloading requires an unlimited data plan or unmetered connection; the model is not downloaded over a metered connection. Source
The model downloads when a site or feature first needs it on an eligible device. Turning off On-device AI removes the model and prevents future downloads. Enterprise policy can block the download too, and the browser deletes the model if free storage drops below 5GB. Local testing on a developer’s own device requires Edge 158 or later and separate feature activation. Site trial participation, device eligibility and a completed model download are different readiness checks. Source Source
The documentation says Prompt API model inputs are processed on the same device and are not collected to train AI models. After the initial model download, it can also work offline. The announcement’s hybrid approach lets developers use the on-device model or a configured cloud provider when needed. The scope of local model-input processing should therefore be distinguished from network requests a site makes separately. Source Source
The accompanying Decisions API is at an earlier stage: a developer preview in Edge Canary and Dev, experimental browser channels. It returns probabilities and confidence values for yes/no, multiple-choice or scoring questions about text in a single pass. It supplies values for classification instead of generating a long textual response. Both the API and its implementation may change, so treating it as the same stage as the live-site Prompt API trial would obscure its availability. Source
OYOPICK’s Take
A shared browser-provided model could reduce the burden of preparing a separate AI runtime for each site. For users, though, the first download, storage needs and device eligibility are part of the experience. Clear availability and download-progress information before a site offers the feature could help more people understand and use local AI’s benefits.
We also think the visibility of the choice between local and cloud processing matters. People could extend a local model’s properties to their understanding of an entire site’s data handling. We hope sites explain where particular requests are processed and present small-model responses or confidence values in a form that people can review.