Microsoft replaces Phi Silica with Aion Instruct, a local model that can hold a file on the PC

The facts

The first column steps back: 2.3 billion parameters, against 3.8 for Phi Silica. The second steps forward: 32,000 tokens of context, against 3,500. Microsoft published on 13 November 2026, under the byline of Connor Al-Joundi, the plan to replace Windows’ on-device language model. Inference stays on the machine. The blog is blunt: more responsive experiences, better use of device resources, familiar workflows kept.

The calendar is dated. A standalone package has been available since October at aka.ms/tryaion, for developers. Aion Instruct is due to start reaching Copilot+ devices on the Windows Insider Programme on 20 November 2026. The retail switch is aimed at late January 2027: on that day, according to Microsoft Learn, Phi Silica goes away. Until then, both models can sit together; the active one is gated by a controlled feature rollout (CFR), and developers can test side by side through a registry key. Microsoft is also expanding Aion to supported GPUs, including NVIDIA RTX-series cards and AMD RDNA4. A technical report is promised, with no date.

The Learn documentation, read on 14 November, adds two clauses that change a deployment. Unlike Phi Silica, Aion Instruct no longer needs an LAF (Limited Access Feature) token; the blog aims at January for that removal. And since October 2026, some Windows AI text APIs insert an invisible watermark to mark generated text. The Aion blog confirms it: invisible watermarking to meet text-provenance requirements in specific regions. Most applications will need no change. Those that rely on highly specific system prompts, custom LoRA adapters, or that parse the output, will have to retest: the model no longer writes like Phi Silica.

Editorial view

A smaller model that moves from 3,500 to 32,000 tokens of context is not a press-release line. It is the difference between summarising one email and holding the thread of a file. For a small business on Windows, the subject is not the name Aion. It is that processing can stay on the PC, including, in time, on a machine without a Copilot+ chip, if the GPU is on the list. The list still has to be there, and the retail model still has to ship. 20 November only opens the Insider channel.

What this changes in practice. Nothing to install tomorrow morning for a freelancer who is not an Insider. The idea can be reused at once: a summary, a sort, a meeting write-up, on the machine, without sending the client file into a US cloud. That is already the logic of the directory’s local agents. The watermark does not wait for January: Learn dates it to October, on some text APIs. That is the gesture our marker on the 2 December deadline describes for labelling generated content.

Watch-outs. Local does not erase the GDPR. If the application still sends the prompt to a cloud service, or if a Microsoft account syncs, the processing is no longer “on the device”. The AI Act does not stop at the NPU either. Learn warns: you remain responsible for labelling if you strip the watermark. On cost: a local model avoids the token bill; it moves the bill onto the machine (GPU, memory, possibly a model download). Learn still describes, for Phi Silica on GPU, a several-gigabyte file downloaded on demand: Aion has, in the sources we read, no size sheet of its own. On lock-in: Aion remains a Windows API. Changing local model will still mean changing Windows. The sources we read give no final GPU list for Aion, no French-language interface, and no date more precise than “late January” for retail.

Trial, or wait. Worth trying, if you already build on the Windows AI APIs: the aka.ms/tryaion package, on a test PC, with no client file. Not worth trying: switching a business app on 20 November onto an Insider channel. Worth waiting for: 20 November for Copilot+ Insiders, late January 2027 for retail, and the promised technical report.

Where it actually helps

Where the file should not leave the desk: let local agents run recurring tasks, sort the inbox once a week, draft a first GDPR processing register, turn meeting notes into a write-up and an action list. 20 November 2026 only opens an Insider channel, on Copilot+ PCs. A write-up pasted into a cloud is not Aion. A write-up reread on the machine, with the watermark still there, is. In late January, Phi Silica will be gone: that is the date to write in the processing register, not the name of the model.