A laptop displays glowing AI workflow icons for chat, coding, decision-making, and cloud synchronization.
Microsoft's Edge team announced three connected on-device AI updates on November 19, 2026. Aion Instruct is now generally available. The Prompt API has an origin trial. The Decisions API has an experimental developer preview. Each has a different maturity level, and that difference matters for what developers should ship.

A laptop displays glowing AI workflow icons for chat, coding, decision-making, and cloud synchronization. What Microsoft announced​

The Edge developer blog lists three items:

  • Aion Instruct reaches general availability as an open-weights model on Hugging Face and as Edge's built-in small language model.
  • The Prompt API gets an origin trial. Developers can use it, along with Aion Instruct, on a live site.
  • The Decisions API arrives as a developer preview. It is an experimental API for fast, structured answers about text.

Microsoft frames this as "hybrid intelligence." Sites use the on-device model when the browser has one and it suits the task. They switch to a configured cloud provider for wider device coverage or heavier inference.

The announcement follows the earlier Aion-1.0-Instruct preview. That preview debuted in Edge Canary and Dev in June. Before it, the Prompt and Writing Assistance APIs had used Phi-4-mini, a 4B-parameter model, for about a year.

Aion Instruct: what it replaces and what we know​

In Edge, Aion now replaces Phi-4-mini as the built-in model. Microsoft puts it at 2.3B parameters. It calls the model smaller, faster, and available on many more devices than Phi, with comparable quality on common web tasks. Those tasks are text generation, summarization, instruction following, and structured outputs.

These are Microsoft's own comparative claims, and the blog post gives no benchmark numbers for the Edge comparison. Microsoft's earlier developer-preview coverage said the model expands support to less capable GPUs and CPU inference.

Edge downloads the model only 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 Edge settings. That removes the model and stops future downloads. Microsoft hasn't published exact hardware thresholds, download size, or the Edge versions that receive the model in the post.

The Hugging Face model card​

The open-weights release is microsoft/Aion-1.0-Instruct-2.3B. Its model card lists these details:

  • An MIT license.
  • Eight supported languages: English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese.
  • A 32,768-token context window and a December 2025 knowledge cutoff.
  • About 2.29B parameters.
  • Loading through Transformers requires trust_remote_code=True, because the model uses custom modeling code.
  • Benchmarks Microsoft reports on the card: MMLU 55.15, GSM8K 74.91, and IFEval 83.90.
  • Microsoft says it is publishing the unquantized model first, with quantized versions to follow.

The card is blunt about limits. A 2.3B model can miss fine details, make misstatements, and fail at complex multi-step reasoning. It can be vulnerable to adversarial attacks, and output quality varies by language. Microsoft says not to use it for automated decisions in critical areas such as medical diagnoses or legal judgments. Treat those cautions as the operating manual, not fine print.

The card also says a further-optimized version, paired with a content-moderation model, is planned to become the Windows OS small language model by the end of January 2027. That is a future Windows plan. It doesn't mean Edge's model already includes the moderation component. The blog says Aion is "coming to Windows later this year," as groundwork for on-device AI across Windows and Copilot+ PCs.

The Prompt API origin trial​

Until now, the Prompt API was a developer preview only in Edge Canary and Dev. The origin trial lets sites call it from production pages in Edge for a broader audience. Microsoft wants feedback on API design, model performance, and cross-browser interoperability. That feedback will feed discussions with other browser vendors as the API is considered for standardization. "Considered" is the operative word. Nothing in the announcement says it has been standardized or adopted by other browsers.

Microsoft Learn's general origin-trial documentation describes the mechanism:

  1. Register your domain on the Edge Origin Trials page and accept the terms. The registration form accepts a domain or subdomain, with an option to cover subdomains. Trailing paths and query parameters are ignored.
  2. Copy the generated token.
  3. Add it to your pages in a <meta http-equiv="origin-trial"> tag, or send it in an Origin-Trial HTTP response header.
  4. Visiting Edge users then get the feature without touching edge://flags.

Some practical rules from that documentation:

  • Feature detection and fallback: Microsoft recommends both, in case the API is missing, the token expires, or the trial ends.
  • Token lifetime: Tokens expire after six weeks by default and can be renewed.
  • Early ends: A trial can end early. The documentation gives a security incident and a needed major redesign as examples.
  • Localhost: Token detection works only on SSL-enabled domains. For localhost, turn on the relevant flag in edge://flags.

These are general Edge origin-trial rules. Check the Prompt API trial's own card for its current expiry and minimum Edge version before relying on it. In one search result, an older Prompt API registration page showed a June 16, 2026 expiry. That page is stale relative to this announcement, which is a good reminder to check the live registration.

Edge 158's web platform release notes, dated November 19, 2026, also list a separate origin trial. It covers Prompt API sampling parameters: topK and temperature let you tune model behavior per language-model session. It is a distinct item from the broader trial, so don't conflate the two.

The experimental Decisions API​

The Decisions API is powered by Aion Instruct and is available only in Edge Canary and Dev. Microsoft says to expect both the API and its implementation to change.

It takes text and a set of questions. Each question can be yes/no, multiple choice, or a score. It returns probabilities and a confidence value, computed in a single pass with no text generation. Microsoft's suggested uses are:

  • Inferring user intent and acting on it.
  • Sorting or routing user-generated content.
  • Scoring or ranking items in real time.

The sample creates a DecisionModel with a context string ("Customer reviews for an online electronics store") and three questions:

  • A choice question on whether the review is about quality, shipping, or price.
  • A score question for satisfaction from 1 to 5.
  • A boolean question on whether the customer is asking for help.

One model.decide() call then evaluates the review "Love the headphones, but the box arrived crushed. Can I get a replacement?" The sample shows the shape of the API. It doesn't show output values, latency, accuracy, or how to set confidence thresholds.

Why this matters, and the caveats​

Classification without generation is a sensible fit for a small model. Picking a label is easier than writing a paragraph, and one pass should be cheaper than a generate-and-parse loop. That reasoning is general industry knowledge, not something Microsoft measured in this post. The model card's warnings still apply. A confidence value isn't proof of correctness, and a customer-support router that misreads sarcasm in German is a plausible failure. Microsoft itself asks whether "latency requirements warrant a smaller model."

Practical guidance for developers​

  • Keep the maturity levels separate. Aion is GA. The Prompt API is an origin trial. The Decisions API is a Canary/Dev experiment.
  • Detect, then fall back. On-device availability depends on the browser, the device, and whether the model has downloaded.
  • Test on your target hardware and languages. The model card says quality varies by language.
  • Don't automate high-stakes decisions with model output, as the card advises.
  • Plan for user opt-out. Users can disable on-device AI in Edge settings, which removes the model.
  • Share feedback on GitHub. That is where Microsoft is collecting input on API design and interoperability.

What it means for Windows and IT admins​

For Windows users, the Edge side is mostly invisible. A site can trigger a model download only on capable devices, and users can switch it off. The Windows side is more notable. The model card ties Aion to the planned Windows OS small language model, so Edge is acting as a proving ground for it. IT administrators should watch for Edge policy and documentation updates as the feature spreads. This announcement doesn't describe any enterprise controls beyond the user setting.

Microsoft is betting that small, local models can carry routine web AI tasks, with the cloud as a backstop. The origin trial is the first real test of that bet.

 

References

  1. Hybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edge Microsoft - Windows Blogs 2026-11-19T23:08:46+00:00
  2. Microsoft Edge Origin Trials developer.microsoft.com
  3. Test experimental APIs and features by using origin trials - Microsoft Edge Developer documentation | Microsoft Learn learn.microsoft.com