Microsoft Will Replace Phi Silica With Aion Instruct in Windows 11 This Fall

Microsoft is preparing a significant upgrade to the AI foundation built into Windows 11, replacing the Phi Silica small language model with Aion Instruct as the engine behind two of its most important on-device Windows AI APIs. The transition, scheduled to conclude in late November 2026, represents one of the largest changes yet to Microsoft's local AI platform.
According to a Microsoft developer transition plan reviewed by WindowsForum.com, the migration will occur in three stages: a standalone testing package beginning October 1, rollout to Windows Insiders on October 23, and retail deployment starting November 24, 2026, when Phi Silica will be removed from supported systems.
The transition plan was communicated by Connor Al-Joundi of Microsoft's Windows AI Platform and Tools team under the subject "Windows AI APIs: Phi Silica to Aion Instruct transition plan." The email originated from Microsoft infrastructure and successfully authenticated using SPF, DKIM, and DMARC.
While Microsoft's public documentation confirms the broader rollout strategy—including an October developer preview, Windows Insider deployment, and November retail rollout—it does not yet publish the specific milestone dates contained in the developer communication reviewed by WindowsForum.

Aion Instruct Becomes the Default Windows AI Model​

The transition centers on two core Windows AI APIs:

• LanguageModel API
• ImageDescription API

These APIs enable Windows applications to perform AI inference locally without bundling or hosting their own large language models.
The LanguageModel API supports workloads such as:

• Text summarization
• Content rewriting
• Prompt-based text generation
• Local natural language processing

Meanwhile, the ImageDescription API generates:

• Accessibility descriptions
• Diagram explanations
• Detailed image summaries
• Visual content descriptions

Microsoft says Aion Instruct will preserve compatibility with the existing API surface, allowing most applications to continue functioning without requiring major code changes.
However, API compatibility does not guarantee identical AI behavior.
Different underlying language models frequently produce different:

• Response formatting
• Prompt interpretation
• Content filtering
• Structured JSON output
• Summarization quality
• Latency characteristics

Applications that simply display AI output may require little more than validation testing. Applications that parse generated text or depend upon deterministic output should expect more extensive regression testing.

Microsoft's Migration Timeline​

October 1: Standalone Testing Package​

The first stage begins with a sideloadable Aion Instruct testing package.
This package allows developers to evaluate applications before Aion reaches users through Windows Update.
Microsoft recommends developers:

• Validate existing applications
• Test LanguageModel API behavior
• Test ImageDescription API behavior
• Retrain custom LoRA adapters
• Prepare application updates
• Verify correct LoRA selection

One of the more notable platform changes involves Microsoft's Limited Access Feature (LAF) system.
Current Phi Silica applications require developers to obtain a Limited Access Feature token before shipping applications that use the model.
According to Microsoft's migration guidance, Aion Instruct removes this requirement entirely, lowering the barrier for independent developers and applications distributed outside the Microsoft Store.

October 23: Windows Insider Rollout​

The second stage begins on October 23, when Aion Instruct starts deploying to:

• Copilot+ PCs
• Selected GPU-equipped Windows devices
• Windows Insider builds

During this period, Phi Silica will remain installed.
Microsoft will instead rely on its Controlled Feature Rollout (CFR) system to determine which model individual devices receive.
As a result, two systems running the same Insider build may temporarily execute different AI models.
To simplify compatibility testing, Microsoft plans to provide a registry-controlled override allowing developers to switch between Phi Silica and Aion Instruct.
Applications using custom LoRA adapters will also be able to inspect a Windows velocity key to determine which model is active before loading the correct adapter.
Microsoft has not yet published either the registry path or the velocity key identifier.

November 24: Retail Rollout​

Microsoft's current target for general availability is November 24, 2026.
Beginning on that date:

• Aion Instruct becomes the production model
• Phi Silica is removed
• Retail Windows devices begin receiving the transition
• Supported GPU devices require updated graphics drivers

Microsoft has not yet published:

• Final GPU compatibility lists
• Required NVIDIA driver versions
• Required AMD driver versions

Those details are expected before retail deployment and will be particularly important for enterprise administrators managing GPU-enabled Windows AI workloads.

Custom LoRA Models Must Be Retrained​

The most significant compatibility challenge involves LoRA adapters.
LoRA (Low-Rank Adaptation) allows developers to specialize a language model for particular datasets, workflows, industries, or writing styles without retraining the underlying foundation model.
Because each adapter is trained against a specific base model, existing Phi Silica LoRAs cannot be assumed to work correctly with Aion Instruct.
Microsoft is directing developers to:

• Retrain adapters using Foundry Toolkit
• Validate adapter behavior
• Ship both Phi-compatible and Aion-compatible adapters during Insider testing
• Detect the active Windows model before selecting an adapter

Loading the wrong adapter against the wrong foundation model could produce incompatibility errors or unpredictable output.

Understanding Aion Instruct​

Microsoft introduced Aion 1.0 Instruct during Build 2026 as the successor to its current on-device small language model.
The company positions Aion Instruct as:

• Smaller
• Faster
• More efficient
• Better suited for local AI workloads

Target workloads include:

• Summarization
• Rewriting
• Intent recognition
• Accessibility
• Local natural language generation

Executing these workloads directly on supported hardware reduces cloud inference costs while improving responsiveness and enabling offline functionality.
Developers should also distinguish Aion Instruct from Aion 1.0 Plan, Microsoft's larger 14-billion-parameter reasoning model intended for more advanced agentic scenarios. The Windows AI API transition concerns Aion Instruct only.
Current public previews remain limited primarily to ARM64 Snapdragon Copilot+ PCs using Qualcomm NPUs. Intel and AMD x64 support is still listed as forthcoming, and the existing preview does not provide a CPU fallback path.
The October developer package is expected to broaden testing capabilities ahead of general availability.

API Stability Doesn't Eliminate Testing​

Although Microsoft intends to preserve the existing Windows AI API surface, replacing the underlying model still introduces behavioral risk.
Developers should verify changes in areas including:

• Response wording
• Output length
• Prompt interpretation
• JSON formatting
• Table generation
• Content moderation
• Summarization quality
• Image descriptions
• Accessibility metadata
• Time to first token
• Overall inference performance
• Memory utilization
• NPU and GPU acceleration

A representative test suite containing fixed prompts, images, and expected outputs provides the best method for identifying regressions before deployment.
Microsoft also recommends recording diagnostics such as:

• Active AI model
• LoRA version
• Windows build
• Accelerator type
• GPU driver version

Those diagnostics will prove especially valuable while Controlled Feature Rollout creates a mixed deployment environment.

Developers Face a Short Migration Window​

The timeline gives developers roughly three weeks between the planned release of the standalone testing package and the start of the Windows Insider rollout.
Applications that rely solely on the built-in APIs may require only compatibility testing. Applications built around custom LoRAs, structured AI output, or GPU inference face a considerably larger migration effort.
By Microsoft's target retail rollout on November 24, 2026, developers should have completed validation against Aion Instruct, retrained any Phi-based LoRA adapters, and confirmed GPU compatibility where applicable.
The migration delivers a faster model, removes previous access restrictions, and continues Microsoft's strategy of treating built-in AI models as replaceable operating system components. While developers benefit from a stable API surface, the transition also underscores an important reality of modern Windows AI: the intelligence beneath those APIs can evolve independently of the applications built on top of them.