---
title: "Microsoft releases Aion Instruct weights ahead of Windows rollout"
description: "Download Microsoft Aion Instruct in bfloat16 or INT4 ONNX WebGPU, and see how it differs from the planned Windows replacement for Phi Silica."
url: "https://promptszone.com/microsoft-aion-instruct-2-3b-open-weights-windows"
published: "2026-11-19T13:00:05Z"
categories: ["New Models"]
---

# Microsoft releases Aion Instruct weights ahead of Windows rollout

Microsoft’s Aion Instruct 2.3B weights are available in bfloat16 and INT4 ONNX WebGPU. Compare the downloads, published scores and Windows rollout targets.

![Microsoft-reported Aion Instruct benchmark scores for the unquantized model: IFEval 83.90%, GSM8K 74.91% and MMLU-Pro 33.35%.](https://promptszone.com/assets/uploads/c21357f7891188d2016d9d981c8f7e4e.png)

**Summary:**

Aion Instruct 2.3B is downloadable under MIT in two formats. The Windows replacement for Phi Silica is a separate rollout.

Microsoft has published downloadable weights for Aion 1.0 Instruct, a 2.3-billion-parameter text model for local tasks such as rewriting and summarization. The Hugging Face release gives developers two ways to evaluate the model before it becomes part of Windows. The downloadable preview is not the same build that Microsoft plans to ship through the operating system.

## Which Aion download should you use?

Microsoft lists both repositories under the MIT license. The releases share a documented 32,768-token context and support English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese.

| Build | What it contains | When to try it |
| --- | --- | --- |
| [Bfloat16 weights](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B) | Unquantized Safetensors and custom Python modeling code | Evaluate the published unquantized model or experiment with Transformers |
| [INT4 ONNX WebGPU](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B-ONNX-WebGPU) | About 1.3 GB of model files; INT4 weights with FP16 activations and cache | Evaluate a smaller package with a compatible WebGPU runtime |

The bfloat16 quickstart loads custom repository code in addition to model weights. Review that code before running it. The WebGPU model card’s 1.3 GB figure describes the model package, not total memory use while generating.

## What the model card’s benchmarks show

On its unquantized release, Microsoft reports 83.90% on IFEval, 74.91% on GSM8K and 33.35% on MMLU-Pro. IFEval averages two strict instruction-following scores, while GSM8K uses loose answer matching. These scores measure different tasks. If you intend to run the INT4 download, evaluate that build against the prompts and hardware you plan to use rather than assuming the bfloat16 numbers transfer unchanged.

## When does Aion replace Phi Silica in Windows?

Microsoft’s Windows developer post targets Windows Insider availability on Copilot+ PCs in early December and full Windows Retail availability in late January 2027. It also says support will expand to compatible GPUs. These are rollout targets, not a guarantee that every eligible computer will have the model on a particular date. Microsoft plans to pair an optimized Aion build with its content moderation model for Windows.

In its Windows comparison, Microsoft lists 2.3 billion parameters for Aion versus 3.8 billion for Phi Silica, and a 32K input limit versus 3.5K. The larger limit gives apps room for more material in a single request. For developers with a custom Phi Silica prompt, the practical next step is to test representative summaries, rewrites and chat prompts against Aion before the Windows API changes underneath the app.

## Sources

- [Microsoft: bfloat16 files](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B/tree/main) and [model card](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B)
- [Microsoft: INT4 ONNX WebGPU files and details](https://huggingface.co/microsoft/Aion-1.0-Instruct-2.3B-ONNX-WebGPU)
- [Microsoft Foundry on Windows: transition plan](https://devblogs.microsoft.com/foundry-on-windows/aion-instruct-local-ai-windows-gpu/)
