Aion Instruct: Power-User Guide to Microsoft’s Phi Silica Successor

Last updated: November 16, 2026

Aion Instruct is Microsoft’s new on-device language model for Windows and the successor to Phi Silica. It has 2.3 billion parameters and a 32K-token input limit, up from 3.5K. It reaches Windows Insider devices from November 20, 2026, retail PCs in January 2027, and is expanding to NVIDIA and AMD GPUs.

Contents

What is new in Aion Instruct

Microsoft announced the transition on November 13, 2026, in a Windows developer blog post. Aion Instruct takes over the job Phi Silica does today: it is the local language model behind the text features of the Windows AI APIs in the Windows App SDK. Apps keep calling the same APIs, and the model underneath changes.

Attribute Phi Silica Aion Instruct
Parameters 3.8B 2.3B
Input context limit 3.5K tokens 32K tokens
Limited Access Feature token Required Not needed (removal targeted for January 2027)
Custom LoRAs Phi Silica adapters Must be retrained with Foundry Toolkit
Output watermarking None mentioned Invisible watermark for text provenance
Status Removed from retail devices in January 2027 Validation package now; Insider rollout from November 20

The larger context window is the change power users will notice first. A 3.5K-token limit ruled out summarizing long documents or email threads in one pass. At 32K tokens, a local Windows app can send far more text to the model without chunking. Microsoft says a technical report with more detail is coming to its Command Line blog.

Rollout timeline

Per Microsoft’s blog and the Phi Silica documentation, the transition runs in three stages:

  • October 2026: a standalone validation package became available at aka.ms/tryaion for testing and LoRA training.
  • November 20, 2026: rollout to Windows Insider Preview devices begins. Phi Silica stays installed, and a Windows Controlled Feature Rollout decides which model is active. Developers can test both side by side through a registry key.
  • January 2027: Aion Instruct reaches retail devices and Phi Silica is removed.

Microsoft is also presenting the model in session THR888, “Build On-Device AI Experiences w/ Windows AI APIs”, at Ignite 2026, with a recording promised.

Hardware and GPU support

Like Phi Silica, Aion Instruct is designed for Copilot+ PC NPUs and is expanding to supported discrete GPUs, including NVIDIA RTX-series and AMD RDNA4 cards. Microsoft says hardware support will grow in stages as configurations meet its quality bar. It has not published a final GPU list for Aion Instruct.

The current Aion Instruct Preview sample is narrower. Its validation package supports three Copilot+ PC platforms, and there is no CPU fallback:

Platform Execution provider Minimum version
ARM64 Snapdragon QNN 2.2480.49
x64 Intel Lunar Lake OpenVINO 1.8.95
x64 AMD Ryzen AI VitisAI 1.8.80 (framework 1.1.0.2 or newer)

For comparison, Microsoft’s docs list Phi Silica GPU support on NVIDIA GeForce RTX 30 series and newer with 6 GB or more of VRAM, with Developer Mode on and the latest vendor driver. If you want to run larger open models on a consumer GPU instead, see our guides to llama.cpp 1.0 and Edge0-145B local setup.

How to try Aion Instruct now

The sample repository includes a WinUI 3 chat app with Chat, Describe image, Text intelligence and LoRA validation tabs. To run it:

  1. Use a supported Copilot+ PC running Windows 11 (Snapdragon, Intel Lunar Lake or AMD Ryzen AI).
  2. Install the .NET 9 SDK and Git with winget, and turn on Developer Mode under Settings, Privacy & security, For developers.
  3. Clone github.com/microsoft/Aion-Instruct-Preview-Sample and open PowerShell in the folder.
  4. Run Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass, then .\Bootstrap.ps1.
  5. Wait for the first model load. The script installs Windows App Runtime 2 if needed, the model framework and the SDK NuGet, and the first NPU cache compile can take several minutes.

The sample requires SDK NuGet 1.1.1 with framework 1.1.0.0 or newer. Visual Studio 2026 also works; Visual Studio 2022 is not supported for its F5 path.

Calling the API

The preview namespace AionInstructPreview.Text mirrors Microsoft.Windows.AI.Text from Windows App SDK 2.0, but implements only a subset of members. The core pattern is LanguageModel.CreateAsync(), then CreateContext() for multi-turn history, then GenerateResponseAsync(context, prompt, options).

  • Options are required. Passing null is rejected. New LanguageModelOptions default to Temperature 0.9, TopP 0.9 and TopK 40.
  • System prompts. CreateContext(String) sets a system prompt for the whole conversation. A context cannot be changed, so start a new one to change the prompt.
  • Streaming. The Progress callback delivers text chunks, not the full text. Read result.Text only when the status is Complete, because a later moderation check can still block the response.
  • Statuses. Handle PromptLargerThanContext by offering a new conversation, and treat the two content moderation statuses separately from Error.
  • Moderation is mandatory. The default severity threshold is Low, and there is no off switch.
  • Built-in skills. TextSummarizer, TextRewriter and TextToTableConverter cover summaries, rewrites with tones and text-to-table conversion.

Prompting a 2.3B model works best with tight instructions and constrained output. The repository includes a prompt guide for Aion and other small models. Our guide to BotPsy 0.8B shows a similar small-model approach for tool calling.

Migration checklist for Phi Silica apps

  1. Install the validation package from aka.ms/tryaion and run your app’s key scenarios against Aion Instruct.
  2. Rewrite and retest custom system prompts, since output wording and behavior differ from Phi Silica.
  3. Retrain any custom LoRA with Foundry Toolkit. Phi Silica adapters, and arbitrary Hugging Face adapters, are not compatible.
  4. Check any code that parses model output, because of the new invisible watermark.
  5. From November 20, use an Insider build and the registry key to compare both models side by side before the January retail switch.

Microsoft asks developers to file feedback as issues in the microsoft/WindowsAppSDK repository.

Limitations

  • Not for shipping yet. The validation package is for developer testing only and must not be shipped to end users.
  • Preliminary speed. The sample says its latency and throughput figures do not represent the final Windows release.
  • Responsible AI testing. The package has not been fully tested for cross-prompt injection or emotional inference.
  • Region limits. Phi Silica features are not available in China, and Microsoft has not stated a different policy for Aion Instruct.

For other recent small open models you can run locally, see our Alpie Seed 9B guide and Qwen3.8-27B GGUF guide.

Frequently asked questions

What is Microsoft Aion Instruct?

Aion Instruct is Microsoft's new on-device language model for Windows. It replaces Phi Silica behind the Windows AI APIs, with 2.3 billion parameters and a 32K-token input context limit.

When will Aion Instruct reach Windows PCs?

Microsoft plans to start rolling it out to Windows Insider Preview devices on November 20, 2026, and to retail devices in January 2027, when Phi Silica is removed.

Do I need to change my app for Aion Instruct?

Most apps need no changes. Microsoft says apps with highly custom system prompts, custom LoRAs, exact-output parsing or dependence on Phi Silica's specific wording should be validated, and LoRAs must be retrained.

Does Aion Instruct run on a GPU?

Microsoft says Aion Instruct is expanding to supported GPUs, including NVIDIA RTX-series and AMD RDNA4 cards, in stages. The current validation package runs only on Copilot+ PC NPUs, with no CPU fallback.

Do I still need a Limited Access Feature token?

Microsoft says LAF tokens are not needed with Aion Instruct. Its blog targets removal of the token requirement for January 2027.

Why does Aion Instruct output contain a watermark?

Microsoft added invisible watermarking to meet text provenance requirements in some regions, including the EU AI Act. The sample notes the watermark may become visible when output is pasted into certain apps.

Sources

Related answers

Leave a Comment