What just shipped
The package now available is the standalone, sideloadable Aion Instruct build Microsoft promised for early October. It exists for two jobs. First, app validation: developers who ship Windows AI API features today on Phi Silica can install the Aion build and confirm their apps behave identically against the replacement model before it arrives through the update channels. Second, LoRA retraining: anyone who fine-tuned adapters against Phi Silica needs to re-train them for the new model, and Microsoft's tooling for that is the Foundry Toolkit.
The single most welcome line in the whole announcement is the one about what is gone: unlike Phi Silica, Aion Instruct does not require LAF tokens. The Limited Access Feature token apparatus, the per-app entitlements, the sparse-identity plumbing that developers had to wire up just to call a local language model, is not part of the Aion story. That was the single largest source of developer friction in the Phi Silica era, and Microsoft is retiring the concept along with the model.
The timeline, verified
Microsoft's documentation now carries a firm three-stage schedule, and it is worth quoting precisely because earlier coverage, including this forum's July piece on the transition, carried an older version of the dates.
Now, October 2026: the standalone sideloadable package is available for testing and LoRA training. Download it, validate Aion Instruct for your app, re-train your LoRAs with the Foundry Toolkit.
November 2026: Aion Instruct begins rolling out to Windows Insider Preview devices. Phi Silica remains installed during this phase, and the active model is controlled by a Windows Controlled Feature Rollout, with a registry key developers can use to test the two models side by side on the same machine.
January 2027: Aion Instruct rolls out to retail devices, and Phi Silica is removed.
The forum's July coverage had the transition concluding in late November; the current official position pushes retail completion to January. Treat the January date as the plan of record and the November date as the start of the Insider phase, not its end. The correction matters for anyone who built migration plans off the earlier reporting: the runway is roughly two months longer than first announced, but the endpoint is the same, and Phi Silica's removal is no longer conditional language. It is a dated line item.
What Aion Instruct actually is
For readers meeting the name here for the first time: Aion Instruct is Microsoft's next-generation on-device small language model, built by Microsoft Research, positioned as the everyday text-intelligence engine for Windows. Its job description is summarization, rewriting, intent detection, extraction, and accessibility scenarios, the routine language work that should never need a cloud round trip. It succeeds Phi Silica, the 3.3B-parameter model that has powered the Windows AI APIs on Copilot+ PCs, and Microsoft describes it as smaller, faster, and smarter than its predecessor. The exact parameter count has not been disclosed, and Microsoft has published no benchmark numbers for the model, so treat any specific figures seen in secondary coverage as speculation.
It is worth keeping Aion Instruct distinct from its sibling, Aion 1.0 Plan. Plan is the 14-billion-parameter reasoning and tool-calling model with a 32K context window, intended for local agentic workflows, announced at Build 2026 as "coming in the coming months." As of this writing there is no verified general availability for Plan, and it is not part of this announcement. One model at a time.
What changes for developers
The good news is architectural: Aion Instruct arrives as a model swap behind the existing Windows AI APIs, not as a new SDK. The LanguageModel surface, GetReadyState, EnsureReadyAsync, CreateAsync, GenerateResponseAsync, and the Text Intelligence Skills built on top of it, stays the same. For most apps the migration should be closer to a validation exercise than a rewrite, which is exactly why Microsoft is shipping the standalone package now: the whole point is to find the apps where "should" turns out to be wrong.
The practical checklist for the next two months is short. Download the validation package and run your app's AI features against it, especially prompt-compression and speculative-decoding paths if you depend on them, since those are hardware-dependent behaviors that a new model can change. Re-train your LoRA adapters against the Aion weights through the Foundry Toolkit; adapters trained for Phi Silica do not carry over. And when the Insider rollout begins in November, use the registry-key side-by-side mode on your own test machines rather than discovering differences from user reports.
One workflow note that carries over from the Phi Silica GPU era: on non-Copilot+ hardware, model components may still arrive as on-demand downloads of several gigabytes through Windows Update, so the consent-dialog pattern for EnsureReadyAsync remains the responsible UX. The model changes; the manners do not.
Why this matters beyond the model
Step back and the significance is not a benchmark number Microsoft has not published anyway. It is that Windows is about to perform its first in-place replacement of the built-in OS language model, on hundreds of millions of machines, with the old model deleted at the end. Phi Silica's removal in January is a real event: every app, sample, and tutorial that names the model needs to be re-examined, and the LAF token infrastructure that developers cursed for two years goes with it. The forum's own Phi Silica update KBs, the August AMD and Intel component updates, become historical documents the moment the swap completes.
There is also a strategy tell. Aion Instruct is a Microsoft Research model, not an OpenAI one, and every routine inference it handles on-device is one that never touches a metered API. The "unmetered intelligence" framing Nadella has been using all year gets a little more literal with each local model that ships. Whether developers actually adopt it at scale is the open question, and the next two months of validation reports will be the first real data.
The caveats, kept honest
As always with a Microsoft model announcement, the checklist of what is not yet established: no published benchmarks for Aion Instruct against Phi Silica or anything else; no disclosed parameter count or architecture details; no word on whether the July open-weights promise for Hugging Face ever materialized; and hardware coverage that is still best on NPU-equipped Copilot+ PCs, experimental on capable discrete GPUs, with CPU inference proven in Edge but the Windows-side story still filling in. The January retail date is a plan, and Microsoft's local-AI dates have slipped before. The validation package is the right place to start forming opinions, and opinions formed from running the thing beat opinions formed from reading about it. That is, after all, what the package is for.