Windows on-device models will move from Phi Silica to Aion Instruct; the retail rollout is not out yet
On 13 November 2026 Microsoft published a plan to move the on-device language model from Phi Silica to Aion Instruct. Official figures: Aion Instruct 2.3B parameters and a 32K-token input limit; Phi Silica 3.8B and 3.5K tokens. Insider Preview is planned from 20 November 2026; retail is targeted for late January 2027.
This item follows Microsoft’s Windows developer blog post of the same date. It is a staged plan, not a completed retail launch, and it is not a test by 全民AI. Expected descriptions such as faster or more efficient performance are not treated here as measured results.
On 13 November 2026 Microsoft said the on-device language model will transition from Phi Silica to Aion Instruct and expand to supported GPUs, with official examples including NVIDIA RTX-series and AMD RDNA4-generation hardware. The official comparison table lists Aion Instruct at 2.3B parameters and a 32K-token input context limit, and Phi Silica at 3.8B and 3.5K tokens.
The official timeline is: a standalone test package was already available in October 2026 at aka.ms/tryaion; rollout to Windows Insider Preview devices is planned to begin on 20 November 2026; retail devices are targeted for late January 2027 to complete the transition. The introduction also says the late-November start is for CP+ devices through the Windows Insider Program. Those dates are plans. Hardware support will expand in stages.
Microsoft says action is required for apps that use highly custom prompting, custom LoRAs, custom model-output parsing because of the new watermarking, or that rely on Phi Silica’s specific outputs; most general apps are described as needing no special steps. Migration notes are in the Windows AI documentation. No device was tested for this item.
On 13 November 2026 Microsoft said the on-device language model will transition from Phi Silica to Aion Instruct and expand to supported GPUs, with official examples including NVIDIA RTX-series and AMD RDNA4-generation hardware. The official comparison table lists Aion Instruct at 2.3B parameters and a 32K-token input context limit, and Phi Silica at 3.8B and 3.5K tokens.
The official timeline is: a standalone test package was already available in October 2026 at aka.ms/tryaion; rollout to Windows Insider Preview devices is planned to begin on 20 November 2026; retail devices are targeted for late January 2027 to complete the transition. The introduction also says the late-November start is for CP+ devices through the Windows Insider Program. Those dates are plans. Hardware support will expand in stages.
Microsoft says action is required for apps that use highly custom prompting, custom LoRAs, custom model-output parsing because of the new watermarking, or that rely on Phi Silica’s specific outputs; most general apps are described as needing no special steps. Migration notes are in the Windows AI documentation. No device was tested for this item.
Usage conditions
- Hong Kong
- This is a staged Windows platform transition, not a region-specific product launch. Whether a user in Hong Kong can use it depends on whether the device is supported and whether the Insider or later retail rollout reaches that device. It was not tested here.
- Macau
- This is a staged Windows platform transition, not a region-specific product launch. Whether a user in Macau can use it depends on whether the device is supported and whether the Insider or later retail rollout reaches that device. It was not tested here.
- Taiwan
- This is a staged Windows platform transition, not a region-specific product launch. Whether a user in Taiwan can use it depends on whether the device is supported and whether the Insider or later retail rollout reaches that device. It was not tested here.
Official source: Microsoft Aion Instruct: Advancing local AI on Windows and expanding to GPU