AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Open Source

Microsoft quietly gives away the AI model it is building into Windows, four months later than promised

Microsoft has made the small AI model it is building into Windows free for any company to use, change and sell. The model, Aion 1.0 Instruct, appeared on the Hugging Face download site on November 18 under the MIT license, one of the least restrictive licenses in software.

Neither OrcaRouter, an AI model marketplace that tracked the release, nor machine.net could find any launch post from Microsoft. The model also arrived roughly four months after the date Microsoft had set. At its Build developer conference in June, the company said the free download would come in July.

For businesses, the timing matters more than the delay. Microsoft's own documentation says Aion Instruct will replace Phi Silica, the model behind Windows' current on-device AI features, on retail PCs in January 2027. Apps that use Windows' built-in language model will then run on this model.

What Microsoft is giving away, and why it matters

Aion Instruct is a small language model: a compact AI that runs on the PC itself rather than in a data center. That means no per-use cloud fees and no data leaving the machine. Microsoft pitches it for summarizing, rewriting and answering questions about documents.

The new model is smaller than the one it replaces, at 2.3 billion parameters versus Phi Silica's 3.8 billion. Parameters are a rough measure of a model's size. It can also take in far more text at once, about 32,000 tokens against 3,500, or roughly a long report instead of a few pages.

Two changes widen who can use it. Phi Silica was a restricted feature: developers had to request an unlock token from Microsoft to call it. Microsoft says Aion Instruct needs no such token, and the requirement is due to go away in January.

Phi Silica also ran only on the dedicated AI chips in Copilot+ PCs. Aion Instruct is expanding to ordinary graphics cards, starting with Nvidia RTX 30-series and newer and recent AMD Radeon cards with at least 6 GB of memory. Microsoft says the AI chips still give the best performance.

The MIT license is the bigger signal. Its only real condition is keeping Microsoft's copyright notice. Aion Instruct carries no revenue limit, no user cap and no country exclusions. A software maker could ship it inside a Mac or Linux product, or retrain it, without asking Microsoft.

That makes the model a low-cost way for Microsoft to court developers. Apps written against Windows' AI features can be tested and adjusted on the open copy, then run on any supported PC where Windows supplies the model at no charge.

What nobody has checked yet

Microsoft's performance figures are its own. The model card lists scores such as 55 on MMLU, a broad general-knowledge test, and 84 on IFEval, which measures how well a model follows instructions. It gives no comparison with rival models, or even with Phi Silica.

As of November 19, no independent tester had published results, according to OrcaRouter. The independent rating service Artificial Analysis had not yet listed it. Safety results appear only as a chart image, with no numbers in the text.

Microsoft itself is modest about the model. The card warns that it may struggle with multi-step reasoning, varies in quality across languages and was "not trained to excel at world knowledge." A promised technical report has no publication date.

The download also is not quite what will ship. Microsoft says a further-optimized version, paired with a separate content-filtering model, will become the Windows model by the end of January. Uptake so far is thin. The repository showed almost no downloads in its first day, and ModelFit, which tests models on consumer hardware, notes that popular run-it-yourself tools such as Ollama and LM Studio cannot yet load it.

What changes for companies building on Windows

Microsoft says most apps using Windows' AI features should keep working without changes when Aion Instruct takes over. But it tells developers to retest apps that rely on carefully tuned instructions or that read Phi Silica's exact wording.

Companies that customized Phi Silica face rework. Custom add-on training made for the old model, known as LoRA adapters, will not carry over and must be redone, according to Microsoft's guidance.

Microsoft has also added an invisible watermark to some AI-generated text from these Windows features, for what it calls provenance rules in certain regions. It says app makers remain responsible for meeting any AI-labeling rules that apply to them, and warns that apps which process the model's output may need changes.

Microsoft's rollout dates are not fully settled. A November 13 company blog post put the first preview release on test builds of Windows in early December. Its developer documentation says November. Both point to January for ordinary PCs.

The larger Aion 1.0 Plan, a 14-billion-parameter model Microsoft described in June as able to plan tasks and use other software, still has no public ship date. Microsoft has not said whether that model will also be free to download.