# Microsoft releases Aion 1.0 Instruct weights, its 2.3B Windows model

> The MIT-licensed small model handles on-device text tasks, and Microsoft says an optimized version becomes the built-in Windows model by January 2027.

URL: https://claudegpt.si/news/microsoft-aion-1-0-instruct-open-weights  
Published: 2026-11-19T10:57:15.243Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Models  
Topics: Microsoft, Small models, Open-weight models, Hugging Face  
Publisher: ClaudeGPT Si

## Key takeaways

- Microsoft published Aion 1.0 Instruct, a 2.3B-parameter on-device model, on Hugging Face on November 18, 2026 under the MIT license.
- It reads up to 32,768 tokens and supports eight languages.
- Microsoft says an optimized version will become the built-in Windows small model by the end of January 2027.

Microsoft published the weights of Aion 1.0 Instruct, a 2.3-billion-parameter language model built to run on Windows PCs, on [Hugging Face](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B) on November 18, 2026, under the MIT license. The model card calls Microsoft Aion Instruct "the foundational Small Language Model (SLM) in the Windows ecosystem" and says a further optimized version will become the small model built into Windows by the end of January 2027.

That makes it the model Windows apps will reach for when they summarize, rewrite or sort text on the device without sending it to the cloud. Microsoft [first announced Aion 1.0 at Build in June](https://blogs.windows.com/windowsdeveloper/2026/06/02/build-2026-furthering-windows-as-the-trusted-platform-for-development/), describing Instruct as "smaller, faster and more efficient than our current Windows OS SLM" and promising open weights on Hugging Face in July. They arrived four months late.

The card lists a 32,768-token context window, eight languages (English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese) and a hybrid design in which 21 of its 28 layers look only at a 2,048-token sliding window rather than the whole input. Microsoft's own scores include 55.15 on MMLU, 74.91 on GSM8K grade-school math and 83.90 on IFEval, a test of following instructions. The card gives no scores for other models to compare against. Microsoft says it is publishing the full-precision model first, with quantized versions to follow, and points to a [sample repository](https://github.com/microsoft/Aion-Instruct-Preview-Sample) on GitHub.

Its bigger sibling, Aion 1.0 Plan, is a 14-billion-parameter reasoning and tool-calling model that Microsoft said at Build would ship inside Windows on capable devices. Its weights are not part of this release. For Microsoft's cloud models announced this week, see our article on [MAI Code 2](https://claudegpt.si/news/mai-code-2-microsoft-ai-models).

## Sources

- [Aion-1.0-Instruct-2.3B model card](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B), Microsoft
- [Build 2026: Furthering Windows as the trusted platform for development](https://blogs.windows.com/windowsdeveloper/2026/06/02/build-2026-furthering-windows-as-the-trusted-platform-for-development/), Microsoft
- [Aion-Instruct-Preview-Sample](https://github.com/microsoft/Aion-Instruct-Preview-Sample), Microsoft
