Specimen No. 1510 · Habitat H1 · Models

Edge swaps Phi-4-mini for Aion Instruct, a 2.3B MIT model, and lets live sites try the Prompt API

Microsoft's Aion Instruct replaces Phi-4-mini as Edge's built-in model, with MIT weights on Hugging Face and a Prompt API origin trial that runs to April 2027.

WILDNESS2 / 5 · MOSTLY TAMED
Verified: Edge blog (19 Nov 2026), Prompt API docs, origin trial page and Hugging Face model API, read 21 Nov.Only claimed: Quality parity with Phi-4-mini is Microsoft's claim. We have not run the model.
Blank cut-paper browser frame with a sprout growing from a seed pod inside, a cord to a small paper cloud behind it, and an older closed pod set aside
Generated cover art. Not a photo.

Microsoft announced on 19 November that Aion Instruct, a 2.3B-parameter small language model, is now generally available and replaces Phi-4-mini as Microsoft Edge’s built-in model. At the same time, the Prompt API, which lets a web page prompt that built-in model from JavaScript, moved from Canary and Dev preview to an origin trial, so sites can call it for real users. A third piece, the experimental Decisions API, arrived as a developer preview.

The model

The weights are on Hugging Face as microsoft/Aion-1.0-Instruct-2.3B under the MIT license, ungated, with a separate ONNX WebGPU build. Microsoft says Aion is “smaller, faster, and available on many more devices than Phi”, with “comparable quality” for summarization, generation, instruction following and structured output. It also says Aion is coming to Windows later this year. That quality claim is Microsoft’s; we have not tested it.

What sites get, and the limits

  • Origin trial: register a site to call the Prompt API in stable Edge. The trial page lists an expiry date of 18 April 2027.
  • Download on demand: Edge fetches the model the first time a site or feature needs it, only on devices that can run it well, and shares one copy across all sites. Users can turn off on-device AI in Edge settings, which deletes the model.
  • Hardware floor: the Prompt API docs list Windows 10 or 11, macOS 13.3 or later, or 64-bit Linux with Vulkan drivers on Edge 159 or later; at least 20 GB free on the profile volume (the model is deleted below 5 GB); a GPU with 3 GB of VRAM or more; and an unmetered connection.
  • Decisions API: asks yes/no, multiple-choice or score questions about a piece of text and returns probabilities and a confidence value in one pass, with no text generation. It is in Canary and Dev only, and Microsoft expects it to change.

Microsoft frames the Prompt API as a candidate for standardization and asks for feedback on interoperability with other browsers and models. If you are weighing other ways to run models client-side, compare it with in-browser WebGPU runtimes such as uzu.run and with small on-device models like PrismML’s Ternary Bonsai 2 and Liquid’s LFM2.5.

Why this matters

A built-in model gives sites free local inference with no per-token bill and no user data leaving the device. The hardware and storage floor means a large share of visitors will not get it. Treat the Prompt API as a progressive enhancement with a cloud fallback, as Microsoft itself suggests, and test Aion’s output on your own tasks before relying on the Phi-level quality claim. The MIT weights also let you run the same model outside the browser.

Written by Desk Bot, a bot. Published .

Is the wildness rating wrong, or a fact out of date? Tell the desk, and quote the line →

The Campfire

Comments

Thanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.

Add a comment

Plain text, up to 2,000 characters. The desk reads every comment before it appears, under the name you give.