Microsoft has published Aion 1.0 Instruct, a 2.3 billion parameter model from its Windows Applied Sciences Group, as open weights on Hugging Face under the MIT license.
The model card says a further optimized version, paired with Microsoft’s content moderation model, “will become the Windows OS SLM by the end of January 2027”.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on models and the harnesses around them. It teaches the general idea only and does not discuss Aion or Microsoft.
The key insight: Microsoft published the open weights of Aion 1.0 Instruct, a 2.3-billion-parameter MIT-licensed model that its card says will become Windows’ built-in language model by the end of January 2027.
What Microsoft Released
Hugging Face’s records show the repository was created on 18 November. Microsoft says it is publishing the unquantized model first, with quantized versions to follow.
The card describes Aion as “the foundational Small Language Model (SLM) in the Windows ecosystem”, built for on-device text work without a constant cloud connection.
On 19 November the Microsoft Edge team followed with a post titled “Hybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edge”. We could not read the post itself.

The Specs
The card lists about 2.29 billion parameters across 28 layers, a 32,768-token context window, a 200,029-token vocabulary and a December 2025 knowledge cutoff.
Of the 28 layers, 21 use sliding-window attention and 7 use full attention, so 75% of the layers look only at a local window of text (our arithmetic on the card).
It supports eight languages, including English, Chinese and Japanese. Multilingual data is about 6% of its training data, which extends the data behind Phi Silica, the previous Windows model.
What Microsoft Claims
Microsoft’s own evaluation table gives Aion 55.15 on MMLU, 74.91 on GSM8K, 48.00 on MATH-500, 83.90 on IFEval and 28.18 on GPQA. These are Microsoft’s scores; we have not run the model.
The card is blunt about the limits. Small models like Aion “can miss fine details, provide misstatements, or show inability to follow complex multi-step reasoning”, it says, and the model “was not trained to excel at world knowledge”.
The Number
Tencent’s Hy4, which we covered this morning, lists 770 billion total parameters. That is about 336 times Aion’s 2.29 billion, and Hy4’s 49 billion active parameters are about 21 times Aion’s whole model (our arithmetic).

The two open models sit at opposite ends of the scale: Aion is built to run on a laptop without the cloud, while Tencent serves Hy4’s FP8 version across eight GPUs, per its card.
Why It Matters Now
The card frames the case for a small local model in its own words: on-device AI that is “responsive, private, and cost-effective” and works “without requiring constant cloud connectivity”.
For related coverage, see our pieces on Apple’s agent job postings and Microsoft’s MAI Code 2 models.
The Structural Read
As we read the card, Microsoft is publishing the preview to collect feedback before the Windows version ships with a content moderation model attached.
As we read the release, the open MIT license lets developers build on the same small model Windows will ship, which suits on-device agents.
Aion 1.0 Instruct model card
“will become the Windows OS SLM by the end of January 2027”
Three Implications
FOR DEVELOPERS The 2.3B weights are downloadable under MIT, with PyTorch, ONNX and GGUF listed and quantized versions to follow.
FOR WINDOWS USERS A further optimized Aion with a content moderation model becomes the Windows OS small language model by the end of January 2027, per the card.
FOR BUYERS COMPARING MODELS Microsoft’s own table lists 83.90 on IFEval and 28.18 on GPQA; the card says the model was not trained to excel at world knowledge.
What Is Not Established
We could not read the Edge post, so we do not report how Edge uses Aion beyond its title. Benchmarks are Microsoft’s own. The Windows version will differ from this preview, the card says.
The Bottom Line
Microsoft has opened the weights of the 2.3-billion-parameter model it plans to make Windows’ built-in language model by the end of January 2027. It is small by design, and the card says so.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Parameters, architecture, context length, languages, training data, benchmark scores and limitation quotes are from Microsoft’s Aion 1.0 Instruct model card; the creation date is from Hugging Face’s records. We could read only the title of Microsoft’s Edge blog post. Benchmarks are Microsoft-reported. The 75% layer share and the Hy4 size comparisons are our arithmetic.
Sources: Aion 1.0 Instruct model card, Hugging Face (microsoft/Aion-1.0-Instruct-2.3B), retrieved 20 Nov 2026 · Microsoft Edge Blog, 19 Nov 2026 (title only; body not readable to us) · MLQ.ai, 20 Nov 2026 · TestingCatalog daily AI brief, 20 Nov 2026 (passed by our editor) · microsoft/Aion-Instruct-Preview-Sample GitHub repository









