Microsoft · Released, not on our provider
Aion 1.0 (2026): what Microsoft’s on-device Windows model is, its benchmarks, hardware and rollout
Aion 1.0 is Microsoft’s on-device small language model for Windows, a 2.3B-parameter model with a 32,768-token context called Aion-1.0-Instruct, published as MIT-licensed weights on Hugging Face and planned to replace Phi Silica as the model built into Windows in January 2027.
Use it now
Chat
This chat starts with Qwen3.8 Flash, not Aion 1.0.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- Microsoft
- Type
- Text
- Released
- 2026-06-02
- OpenRouter id
- None
Checked
Aion 1.0: facts, with sources
- Microsoft’s Edge blog announced a developer preview of the pre-release Aion-1.0-Instruct on June 2, 2026, in Edge Canary and Dev, with an open-source release on Hugging Face planned for July. The weights are public now at microsoft/Aion-1.0-Instruct-2.3B; the model card gives no release date.Source: blogs.windows.com · checked 2026-11-19
- The Hugging Face card calls it the "foundational Small Language Model (SLM) in the Windows ecosystem", optimized for on-device AI, text and operating system use cases. It has about 2.29B parameters ("2.3B"), 28 layers (21 sliding-window layers with a 2,048-token window and 7 full-attention layers), a 32,768-token context, a 200,029-token vocabulary, bfloat16 weights and a knowledge cutoff of December 2025; languages are English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese, and it supports tool calling.Source: huggingface.co · checked 2026-11-19
- The Hugging Face card says it works "entirely on-device, ensuring user data remains private", and Edge’s Prompt API docs say input never leaves the device and is not collected to train AI models; the weights are public under the MIT license and load with trust_remote_code.Source: huggingface.co · checked 2026-11-19
- Microsoft’s Hugging Face collection "Aion" lists two repositories, Aion-1.0-Instruct-2.3B and its ONNX WebGPU build; we found no Microsoft page for a larger "Aion-1.0-Plan" model, which some third-party blogs describe.Source: huggingface.co · checked 2026-11-19
- Microsoft’s Aion is not on OpenRouter’s model list (checked November 18, 2026): the aion-labs/ ids there (such as aion-labs/aion-3.5) come from a different maker, so this site cannot run it. This site is a text chat with other models, such as Qwen3.8 Flash and Inkling 2 Small.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
Aion 1.0 benchmarks: Microsoft’s own numbers for Aion-1.0-Instruct
- Microsoft’s own figures on the Hugging Face card (not tested by us): MMLU 55.15, MMLU-Pro 33.35, MMMLU 45.51, GPQA 28.18, GSM8K 74.91, GSMPlus 49.56, MATH-500 48.00, IFEval 83.90, IFBench 32.00, Multi-IF 64.70 and translation 43.18 chrF++, run with lm-evaluation-harness (GPT-4o as judge where one is needed).Source: huggingface.co · checked 2026-11-19
Vendor benchmarks
Aion 1.0 (Aion-1.0-Instruct) benchmark scores from its Hugging Face card
People who search for Aion 1.0 want to know what a 2.3B model that ships inside Windows can actually do. Microsoft’s card shows a small model: 74.91 on GSM8K grade-school maths and 83.90 on IFEval instruction following, but 28.18 on GPQA and 33.35 on MMLU-Pro, so it is built for summarizing, rewriting and tool calls on your own machine, not for hard reasoning. There is no number against another model on the card, so we do not rank it. You cannot chat with Aion here, but small models you can chat with include Qwen3.8 Flash, Inkling 2 Small and Hy4 Flash.
| Benchmark | Aion-1.0-Instruct |
|---|---|
| MMLU | 55.15 |
| MMLU-Pro | 33.35 |
| MMMLU (multilingual MMLU) | 45.51 |
| GPQA | 28.18 |
| GSM8K | 74.91 |
| GSMPlus | 49.56 |
| MATH-500 | 48.00 |
| IFEval | 83.90 |
| IFBench | 32.00 |
| Multi-IF | 64.70 |
| Translation (chrF++) | 43.18 |
Our table
Aion 1.0 on Windows, in Edge and on Hugging Face: status and what each needs
- Microsoft Learn’s Phi Silica page says Aion Instruct begins rolling out to Windows Insider Preview devices in December 2026, and to retail devices in January 2027, when Phi Silica is removed; LAF tokens are no longer needed, and Phi Silica LoRAs are not compatible (re-train them with Foundry Toolkit).Source: learn.microsoft.com · checked 2026-11-19
- The Hugging Face card says a further optimized version, coupled with Microsoft’s content moderation model, will become the Windows OS SLM by the end of January 2027, and that the public weights are "an introduction to what’s coming".Source: huggingface.co · checked 2026-11-19
- Edge’s Prompt API docs list Windows 10 or 11, macOS 13.3+ or x64 Linux with Vulkan drivers, at least 20 GB free on the Edge profile volume (the model is deleted below 5 GB), a GPU with 3 GB of VRAM or more, a non-metered connection for the download, and a Device performance class of Medium or higher on edge://on-device-internals. The docs call it "a capable 2.3B parameter small language model".Source: learn.microsoft.com · checked 2026-11-19
- Microsoft’s Edge blog says the new on-device models also run on less capable GPUs and, through CPU inference, on devices without a GPU; the Prompt API docs list a 3 GB GPU, and we could not reconcile the two statements.Source: blogs.windows.com · checked 2026-11-19
- Microsoft’s Aion-Instruct-Preview-Sample on GitHub is for developer validation only ("not for end-user or customer use"), needs a Windows 11 Copilot+ PC with an NPU (Snapdragon through QNN, Intel Lunar Lake through OpenVINO, AMD Ryzen AI through VitisAI) and has no CPU fallback; it ships under a pre-release EULA, and the repo says preview performance is not representative of the final Windows release.Source: github.com · checked 2026-11-19
Where it ships
Where Aion 1.0 is available, its status and what it needs
Aion 1.0 reaches people in four ways, at different stages, so "is it out?" has several answers. Microsoft’s Learn page puts the Windows rollout in December 2026 for Insiders and January 2027 for retail devices; in Edge it is a developer preview; the Hugging Face weights are public today; and the GitHub sample is only for developers with a Copilot+ PC. Each row below names what you need. If you only want a small model to chat with today, Qwen3.8 Flash and Inkling 2 Small run on this site, and Hy4 Flash has its own table.
| Where | Status | What it needs |
|---|---|---|
| Windows (built in) | Insider Preview devices from December 2026; retail devices from January 2027, replacing Phi Silica | No LAF token; Phi Silica LoRAs must be re-trained; hardware list not given on the page |
| Edge (Prompt API) | Developer preview behind the Prompt API flag or its origin trial (Canary, Dev and Beta) | 20 GB free disk, a GPU with 3 GB of VRAM, Device performance class Medium or higher, a non-metered download |
| Hugging Face weights | Public under the MIT license; unquantized first, quantized versions to follow | trust_remote_code=True (the repository ships custom code) |
| Windows developer sample (GitHub) | Developer validation only, under a pre-release EULA | A Windows 11 Copilot+ PC with an NPU (Snapdragon, Intel Lunar Lake or AMD Ryzen AI); no CPU fallback |
Source: learn.microsoft.comSource: learn.microsoft.comSource: huggingface.coSource: github.com
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to Aion 1.0 you can use now
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
Inkling 2 Small
Thinking Machines Lab$0.30 / $1 (per million tokens)
- Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
- A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
- Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14
FAQ
Aion 1.0 FAQ
What is Aion 1.0?
Aion 1.0 is a small language model that Microsoft describes as the "foundational Small Language Model (SLM) in the Windows ecosystem", built to run on your own device for summarizing, rewriting, chat and tool calling. The public model is Aion-1.0-Instruct, with about 2.3B parameters and a 32,768-token context. It first appeared as a developer preview in Edge on June 2, 2026.
Is there an Aion 1.0 Plan model?
We could not confirm one. Microsoft’s Hugging Face collection lists only Aion-1.0-Instruct-2.3B and its ONNX WebGPU build, and none of Microsoft’s pages we read mentions a larger model called Plan, although some third-party blogs describe a 14B "Aion 1.0 Plan". Treat that as unverified until Microsoft publishes it.
When does Aion 1.0 come to Windows?
Microsoft Learn says Aion Instruct begins rolling out to Windows Insider Preview devices in December 2026 and to retail devices in January 2027, when Phi Silica is removed. The Hugging Face card says the optimized version, paired with Microsoft’s content moderation model, becomes the Windows OS model by the end of January 2027. These are plans and can change.
What hardware does Aion 1.0 need?
It depends on the route. In Edge, Microsoft’s docs ask for 20 GB of free disk, a GPU with 3 GB of VRAM and a Device performance class of Medium or higher, although the Edge blog also says it can run on the CPU. Microsoft’s GitHub sample for Windows needs a Copilot+ PC with an NPU and has no CPU fallback. The Windows retail hardware list is not on the Learn page.
How good is Aion 1.0?
By Microsoft’s own numbers, which we did not run, it scores 74.91 on GSM8K, 83.90 on IFEval and 64.70 on Multi-IF, but 28.18 on GPQA and 33.35 on MMLU-Pro. That is a profile for everyday text tasks on a device, not for hard reasoning. The card shows no comparison with another model. See Qwen3.8 Flash for a small cloud model.
Can I download Aion 1.0 and is it private?
Yes, the weights are on Hugging Face as microsoft/Aion-1.0-Instruct-2.3B under the MIT license, and you load them with trust_remote_code. Microsoft says the model works entirely on-device so user data stays private, and Edge’s docs say input never leaves the device and is not collected to train AI models.
Can I chat with Aion 1.0 here?
No. Microsoft’s Aion is not on OpenRouter’s model list; the aion-labs ids there are from a different maker. You can chat with a small model instead: on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Hy4 Flash is another.
Is this the official Microsoft website?
No. This page is unofficial and not affiliated with or endorsed by Microsoft. Facts link to the pages we read.