top of page

Microsoft Aion Instruct replaces Phi-4-mini in Edge as Prompt API opens

2 hours ago
8 min read
Cover with the Microsoft logo and the title Microsoft Aion Instruct replaces Phi-4-mini in Edge as Prompt API opens, Data Studios original cover

Microsoft has made Aion Instruct the built-in small language model of Microsoft Edge, replacing Phi-4-mini, and has moved the browser's Prompt API from a Canary and Dev developer preview into an origin trial that developers can run on live websites. The company announced both changes on November 19, 2026 on the Microsoft Edge Blog, together with a third item, an experimental Decisions API that uses the same model to answer yes/no, multiple-choice and scoring questions about a piece of text without generating any prose.


Aion Instruct is a 2.3-billion-parameter model built by the Windows Applied Sciences Group. Its weights are published on Hugging Face under the MIT license as Aion-1.0-Instruct-2.3B, with a 32,768-token context window, and Microsoft describes it as smaller and faster than Phi-4-mini while delivering comparable quality for text generation, summarization, instruction following and structured outputs. The same model is also replacing Phi Silica as the on-device language model of Windows, with a rollout to Windows Insider devices planned for December 2026 and to retail Copilot+ PCs in January 2027.


The practical consequence for web developers is concrete. A site that registers for the origin trial can call a local language model from JavaScript on Edge, without sending the user's text to a server, and the pool of devices that can run that model grows because Aion Instruct also runs on less capable GPUs and on CPU-only machines, where Phi-4-mini required at least 5.5 GB of video memory.


··········


⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

MICROSOFT AION INSTRUCT: WHAT THE 2.3B MODEL IS.


According to the official model card, Aion-1.0-Instruct-2.3B is an instruction-tuned causal language model with about 2.29 billion parameters and tied input and output embeddings. It has 28 layers, a hidden size of 2,048 and grouped-query attention with 24 query heads and 8 key/value heads. Microsoft uses a hybrid attention pattern of 21 sliding-window layers, each looking back 2,048 tokens, and 7 full-attention layers, with RoPE positional encoding and YaRN scaling applied to the full-attention layers to reach the 32K-token context. The vocabulary has 200,029 tokens and the knowledge cutoff is December 2025.


The model card lists eight languages: English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese, with multilingual material making up about 6% of the training data. Microsoft does not give a token count for training. It says the data extends the corpus used for Phi Silica, the previous Windows model, with quality-filtered public documents and synthetic data aimed at math, coding, common sense and world knowledge. Weights are available in PyTorch, with ONNX and GGUF formats listed in the repository metadata, and Microsoft separately publishes an INT4 ONNX build for WebGPU of about 1.3 GB.


Microsoft is direct about the limits. The card says a model of this size can miss details, make factual errors and struggle with complex multi-step reasoning, that it is not designed as a source of world knowledge, and that safety results on cross-prompt injection and two other internal safety metrics still need improvement. Its intended uses are document comprehension, content analysis, question answering and processing of multi-page documents and email threads on the device.


........


Benchmark (Microsoft model card)

Aion-1.0-Instruct-2.3B

MMLU

55.15

MMLU-Pro

33.35

MMMLU (multilingual)

45.51

GPQA

28.18

GSM8K

74.91

MATH-500

48.00

IFEval

83.90

IFBench

32.00

Multi-IF

64.70

chrF++ (translation)

43.18


........


The strongest figure is instruction following, with 83.90 on IFEval, which is the capability a browser API built around short tasks and structured outputs relies on most. Knowledge-heavy tests such as MMLU-Pro and GPQA sit far lower, which is consistent with Microsoft's own warning that the model is meant to work on text the user provides rather than answer open factual questions.


··········


AION INSTRUCT VS PHI-4-MINI AND PHI SILICA: SIZE, CONTEXT AND HARDWARE.


Aion Instruct replaces two different Microsoft models at once: Phi-4-mini in the Edge browser and Phi Silica in Windows. The figures below come from Microsoft's own model cards, Edge documentation and the Foundry on Windows blog.


........


Model

Role

Parameters

Context

License

Hardware note

Aion-1.0-Instruct-2.3B

New Edge and Windows model

2.3B

32K tokens

MIT

Runs on weaker GPUs and CPU-only devices

Phi-4-mini

Previous Edge model

3.8B

128K tokens

MIT

Needed 5.5 GB of VRAM or more in Edge

Phi Silica

Previous Windows model

3.8B

3.5K input tokens

Built into Windows

Copilot+ PC NPUs


........


The comparison shows where the trade sits. Against Phi Silica, Aion Instruct is smaller and has a context window roughly nine times longer for input, which is what lets Windows features handle whole documents instead of short snippets. Against Phi-4-mini, Aion Instruct gives up context length (32K instead of 128K tokens) and about 1.5 billion parameters in exchange for lower hardware requirements.


Published benchmark scores for the two Edge models cannot be compared directly. Microsoft's Phi-4-mini card reports MMLU at 67.3, MMLU-Pro at 52.8 and GSM8K at 88.6, but with few-shot and chain-of-thought settings, while the Aion card reports results from EleutherAI's open evaluation suite and, for some tests, a GPT-4o judge. Microsoft's own claim for the browser is limited to "comparable quality" on the tasks Edge exposes, and no head-to-head table on identical settings has been published.


··········


EDGE PROMPT API ORIGIN TRIAL: WHAT CHANGES FOR WEB DEVELOPERS.


Until this announcement, the Prompt API in Edge was a developer preview limited to the Canary and Dev channels and switched on through a flag at edge://flags. The origin trial lets developers register a site and use the API on production pages for real users, so that Microsoft can collect feedback on the API's design, on model performance and on interoperability with other browsers. Microsoft says that feedback will inform the possible standardization of the API, which is discussed in the W3C Web Machine Learning community.


The API itself is the LanguageModel interface already documented for the preview. A page checks whether the model is available or still needs to be downloaded, creates a session, optionally with a system prompt and few-shot examples, and then calls prompt for a full answer or promptStreaming for token-by-token output. Sessions can be cloned, stopped with an AbortController and destroyed to free memory. The following availability check and JSON-schema constraint are adapted from Microsoft's Prompt API documentation.


const availability = await LanguageModel.availability();

if (availability == "downloadable" || availability == "downloading") {
  // The model can be used, but it needs to be downloaded first.
}

const session = await LanguageModel.create();
const response = await session.prompt(reviewText, {
  responseConstraint: schema
});

The responseConstraint option accepts a JSON schema and forces the model's answer into that shape, which in Microsoft's example returns a sentiment label and a confidence score as parseable JSON. For sites, this is the most useful part of the package, because classification, extraction and tagging tasks run on the visitor's device with output that code can consume directly.


Microsoft's Learn page for the Prompt API still describes the flag-based developer preview, with Phi-4-mini as the default model and Aion-1.0-Instruct as a prerelease option from Edge 150.0.4070. Its system requirements remain the reference until the documentation is updated: Windows 10 or 11 or macOS 13.3 or later, at least 20 GB of free space on the volume holding the Edge profile (the model is deleted if free space falls below 10 GB), and an unmetered network connection for the download. The origin trial registration page and updated documentation are linked from the November 19 announcement.


··········


[[ADPLUS_300x250_1]]

AION INSTRUCT IN EDGE: DOWNLOAD, PRIVACY AND HOW USERS TURN IT OFF.


Edge does not ship the model to every browser. Microsoft says the browser downloads Aion Instruct only when a website or a browser feature first asks for it, and only on devices that can run it well. Users can disable on-device AI in Edge settings, which removes the downloaded model and stops future downloads.


This design keeps the cost of the model on devices that actually use it, but it also means a developer cannot assume the model is present. Microsoft's guidance is what it calls hybrid intelligence: a site uses the on-device model when it is available and suitable for the task, and falls back to a cloud provider the site has configured when it is not. In practice, any production integration needs both paths, with the availability check deciding which one runs. Readers who manage Microsoft's AI features on company machines can find the related settings in our guide on how to disable Microsoft Copilot in Windows, Edge and Microsoft 365 apps.


........


Item

Status announced November 19, 2026

Where

Aion Instruct as Edge's built-in model

Generally available, replaces Phi-4-mini

Microsoft Edge

Prompt API

Origin trial on live sites

Microsoft Edge

Decisions API

Experimental developer preview

Edge Canary and Dev

Aion Instruct weights

Open, MIT license

Hugging Face

Aion Instruct in Windows

Insider devices from December 2026, retail in January 2027

Windows and Copilot+ PCs


........


··········


MICROSOFT DECISIONS API: CLASSIFICATION WITHOUT TEXT GENERATION.


The third announcement is the least mature but the most specific to small models. The Decisions API, available as an experimental preview in Edge Canary and Dev, takes a context, a list of questions and an input text, and returns probabilities and a confidence value for yes/no, multiple-choice or scoring questions in a single pass. Microsoft's example creates a model with DecisionModel.create and then calls decide on the text.


Because the API never generates prose, it avoids the slowest part of running a language model on a laptop and returns structured answers by design. Typical uses include routing a support message, flagging a comment for moderation, or scoring how well a product review matches a set of criteria. Microsoft has published an explainer for the API and asks for feedback through its GitHub discussions, which signals that the interface may still change before any wider release.


··········


⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁤

AION INSTRUCT IN WINDOWS: THE PHI SILICA TRANSITION AND DEVELOPER IMPACT.


The browser change is the visible part of a larger switch. In a November 13 post on the Microsoft Foundry on Windows blog, Microsoft said Aion Instruct is replacing Phi Silica as the on-device language model that powers Windows and Microsoft application features. The schedule given there is a standalone package for testing since October 2026, a rollout to Windows Insider devices in early December 2026 with a beta build targeted for December 4, and a rollout to retail devices in January 2027 that completes the transition. The Hugging Face card adds that an optimized version paired with a Microsoft content moderation model is planned to become the Windows OS model by the end of January 2027.


Microsoft is also extending Aion Instruct on Windows to supported discrete GPUs, naming NVIDIA RTX-series and AMD RDNA4 cards, with hardware support added in stages. Developers who built on Phi Silica with highly customized prompts, custom LoRA adapters or parsing that depends on Phi Silica's exact output strings are told to validate their code, because the replacement model will not reproduce the same responses. Microsoft also says the model adds invisible watermarking to meet text provenance requirements in specific regions, and that removal of the LAF token requirement for developers is targeted for January 2027.


Microsoft's recent model work on the cloud side follows a different track, as shown by MAI-Code-2 in GitHub Copilot and the MAI-Realtime-1 speech model, while Aion Instruct is the company's bet for work that stays on the user's machine.


··········


MICROSOFT AION INSTRUCT AND THE EDGE PROMPT API: CONFIRMED FACTS AND OPEN QUESTIONS.


Several facts are established by Microsoft's own pages. Aion Instruct is a 2.3B-parameter, MIT-licensed open-weights model with a 32K-token context; it is now Edge's built-in language model in place of Phi-4-mini; the Prompt API is in an origin trial that permits use on live sites; the Decisions API is an experimental preview in Canary and Dev; and Windows will move from Phi Silica to Aion Instruct between December 2026 and January 2027.


Other points are still open. Microsoft has not published like-for-like benchmarks against Phi-4-mini, inference speed on target hardware, the exact Edge version and end date of the origin trial in the announcement itself, or hardware requirements specific to Aion Instruct beyond support for weaker GPUs and CPU-only devices. Interoperability is also unsettled, since a standardized Prompt API would run on whatever model each browser ships, and a site that relies on a particular quality level would have to test each one. For developers, the immediate value is a free, local model with schema-constrained output that more Edge users can actually run, provided every integration keeps a cloud fallback for devices where the model is missing.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page