EN AR

Microsoft Edge swaps Phi-4-mini for new 2.3B Aion Instruct model

Tamara khadrah

Key Points

  1. Microsoft said Aion Instruct is now Edge's built-in small language model, replacing Phi-4-mini.
  2. The 2.3-billion-parameter model ships as open weights on Hugging Face and reaches Windows later this year.
  3. Edge developers can now call the Prompt API on live production sites through an origin trial.

The latest:

Microsoft Edge has replaced Phi-4-mini with Aion Instruct as its built-in small language model, Microsoft said in a Windows Blogs post dated November 19, 2026. The 2.3-billion-parameter model is also released with open weights on Hugging Face. Alongside it, Microsoft opened an origin trial for the Prompt API and published a developer preview of an experimental Decisions API that returns structured answers about text.

Details:

  • The model: Microsoft said Aion Instruct runs at 2.3 billion parameters and is smaller, faster and supported on more devices than Phi-4-mini, while matching it on common web tasks including text synthesis and generation, summarization, instruction following and structured outputs. The post did not publish benchmark figures behind the comparison.
  • Distribution: The model is generally available as open weights on Hugging Face and, according to Microsoft, is coming to Windows later this year as a foundation for on-device AI across Windows and Copilot+ PCs. No specific release date for the Windows rollout was given.
  • How it downloads: Edge pulls Aion Instruct only when a site or feature first requires it, and only on devices Microsoft judges capable of running it well, the company said. Users can disable on-device AI in Edge settings, which deletes the model and blocks future downloads.
  • Prompt API: The Prompt API was previously limited to developer preview in the Edge Canary and Dev channels. Microsoft said the origin trial now lets developers call it from live production sites and test on-device inference under real-world traffic.
  • Standards angle: Microsoft asked developers for feedback on API design, model performance and interoperability problems across different browsers and models, saying it will feed discussions with other browser vendors and the web community as the Prompt API is weighed for standardization.
  • Decisions API: The experimental Decisions API, powered by Aion Instruct, lets sites pose yes/no, multiple-choice or scoring questions about a text input and receive probabilities plus a confidence value, computed in a single pass with no text generation. It is a developer preview in Canary and Dev only.
  • Named use cases: Microsoft listed detecting user intent and acting on it, sorting and routing user-generated content, and scoring and ranking items in real time, all from client-side JavaScript. The company cautioned that both the Decisions API and its implementation are expected to change.
  • Hybrid framing: Microsoft described the package as hybrid intelligence on the web: sites use browser-resident models where suitable, then switch to a configured cloud provider to widen device coverage or reach more powerful inference. Existing on-device features cover writing assistance, translation, language detection and speech recognition.
  • Open question: Microsoft said it is asking developers testing the Decisions API whether their latency requirements warrant an even smaller model, and is also seeking performance and interoperability feedback on the Summarizer API.

Background:

Aion Instruct was first shown as a developer preview under the name Aion-1.0-Instruct. Microsoft said it revised the model’s training after early-adopter feedback to improve quality and performance for on-device web workloads before this general-availability release.

Between the lines:

The parameter drop to 2.3 billion, paired with Microsoft’s stated device-capability gating, points at reach rather than raw capability as the goal: a model only downloads where it runs well, so a smaller model widens the installed base a site can rely on. The open question on latency suggests that calculation is not finished.

What’s next

Watch the Windows rollout of Aion Instruct before year-end, developer feedback through the Prompt API origin trial, and whether other browser vendors engage on standardizing the Prompt API.

What to read next