Microsoft Aion 1.0 is one of the most significant things to happen to Windows in years — and most people haven’t heard of it yet. Announced at Microsoft Build 2026 on June 2, Aion 1.0 is a family of AI models built directly into Windows 11 that run entirely on your device, without sending a single byte of your data to the cloud.
No subscription. No API key. No internet connection required.
If that sounds like a quiet revolution for how your PC handles AI, that’s because it is.
What Is Microsoft Aion 1.0?
Aion 1.0 is a family of small language models (SLMs) developed by Microsoft Research and designed to run locally on Windows 11 hardware. Unlike cloud-based AI tools that route your requests through remote servers, Aion processes everything on your machine — using your CPU, GPU, or NPU depending on what’s available.
Microsoft describes its goal as delivering “unmetered intelligence” on Windows. The idea is simple: AI features should work the same way your calculator works. Instantly, privately, and without depending on whether your Wi-Fi is fast enough.
Aion 1.0 is not one model — it’s two, each built for a different kind of task.
Aion 1.0 Instruct: The Everyday AI Layer
Aion 1.0 Instruct is the lighter of the two models. It’s designed for the kind of AI tasks that show up constantly in everyday computing — the background intelligence that most users don’t even consciously notice but rely on all the time.
Microsoft describes it as “smaller, faster, and smarter” than the previous Windows on-device model, Phi Silica (based on Phi-4-mini). It replaces that older model and brings noticeably better performance for routine tasks.
What Aion 1.0 Instruct Can Do
Aion 1.0 Instruct handles:
- Text summarization — condensing long documents into key points
- Rewriting and editing — suggesting cleaner alternatives to what you’ve written
- Intent detection — understanding what you’re trying to do in Windows Search or other tools
- Accessibility features — live caption cleaning, real-time text adjustments
- Right-click intelligence — in-context AI actions in File Explorer and other apps
These are the kinds of features you’ll use dozens of times a day without thinking. They’re fast, private, and available offline.
The Hardware Advantage: It Runs on CPU
Here’s the part that makes Aion 1.0 Instruct genuinely different from most other on-device AI stories.
It runs on a standard CPU. No dedicated GPU required. No NPU required.
Every other major on-device AI feature in 2026 comes with a hardware asterisk — Apple’s Foundation Models require Apple Silicon, many Windows Copilot+ features require a specific NPU. Aion 1.0 Instruct has no such restriction. If you’re running Windows 11 on a modern CPU, the model works on your device.
That’s a massive expansion of the audience compared to anything Microsoft has done with on-device AI before.
Availability
Aion 1.0 Instruct is available now in developer preview through Microsoft Edge Insider channels. Open weights are scheduled to land on Hugging Face in July 2026, which will let developers fine-tune the model for specialized domains and use it outside of Windows entirely.
Aion 1.0 Plan: The Local AI Agent
Aion 1.0 Plan is where things get genuinely ambitious. This is not a helper model for summarization — it’s a full reasoning and tool-calling model designed to run complex, multi-step agentic workflows entirely on your PC.
Key Specs
- 14 billion parameters
- 32K context window
- Reasoning and tool-calling capability
- Designed to orchestrate sub-agents
- Ships in-box with Windows on capable devices
A 14-billion-parameter model running locally on a laptop — without the cloud — was practically unthinkable just two years ago. The combination of more efficient model architectures and hardware like Nvidia’s RTX Spark (which supports up to 128GB of unified memory) makes it possible today.
What Aion 1.0 Plan Can Do
Aion 1.0 Plan is built to act, not just respond. It can:
- Reason over your intent — understand complex, multi-step goals stated in natural language
- Call tools and APIs — interact with Windows system functions, files, and applications
- Manage files — read, organize, and act on documents on your local storage
- Orchestrate sub-agents — delegate parts of a task to other AI processes running locally
- Execute agentic workflows — chain together multiple steps to complete a goal without you clicking through each one
Think of it less like a chatbot and more like a local AI that can work through a task while you focus on something else.
Why Local Matters for Agentic AI
Agentic AI — AI that plans and acts rather than just answers — has a fundamental problem when it runs in the cloud: every planning step, every tool call, and every evaluation costs tokens and round-trips to a remote server. That adds up fast.
When Aion 1.0 Plan runs on your device:
- Sensitive data never leaves your machine. You can reason over confidential documents, financial records, or private files without them touching a third-party server.
- Latency disappears. The planning loop runs at local speed, not network speed.
- Cost drops off. Orchestration overhead — often the bulk of an agentic task’s token consumption — no longer runs through a metered cloud API.
These are not theoretical advantages. They address the exact friction points that have made cloud-based agents feel slow and unreliable in practice.
How Aion Fits Into Windows: The Bigger Picture
Aion 1.0 is the software layer of a much larger platform shift Microsoft is building. The hardware counterpart is Nvidia’s RTX Spark superchip, announced at Computex 2026. Together, they form Microsoft and Nvidia’s vision for what they’re calling “agentic computing” — a PC that doesn’t just run software you operate, but one that can work on goals you set.
The AI Routing Stack
Aion introduces a tiered approach to AI on Windows:
- Aion 1.0 Instruct → handles everyday text tasks locally on any capable CPU
- Aion 1.0 Plan → handles complex reasoning and agent workflows on capable hardware
- Cloud models (GPT, Claude, others) → handle frontier-scale tasks that need more capability
The routing between these tiers is handled through Nvidia’s OpenShell platform, which applies policy-based rules: sensitive data routes to local Aion, general queries can go to cloud models. This is what makes a hybrid architecture practical rather than just theoretical.
Windows AI APIs Expand to More Hardware
Alongside Aion 1.0, Microsoft also expanded Windows AI APIs at Build 2026:
- Speech-to-text recognition now runs on CPUs and NPUs — no dedicated hardware required
- On-device SLM now supports capable discrete GPUs, not just NPUs
- Video Super Resolution now runs on CPUs, making AI video upscaling available on far more machines
The API design abstracts hardware from developers. An app calls a speech-to-text API, and Windows automatically routes the workload to whatever accelerator is available — NPU first, GPU second, CPU fallback. Developers don’t need to write hardware-specific code.
Why Microsoft Built Aion Itself (Not with OpenAI)
This is worth addressing directly. Aion 1.0 is built by Microsoft Research, not by OpenAI.
Microsoft has a deep and well-publicized partnership with OpenAI, which powers much of Copilot. But Aion represents a deliberate move to build in-house models for on-device use — reducing Microsoft’s dependence on OpenAI for a key layer of the Windows platform and lowering costs for both Microsoft and developers.
Every inference that routes through Aion instead of an OpenAI API call saves money. At the scale of hundreds of millions of Windows machines, that’s a significant shift.
The open-weights commitment — releasing Aion 1.0 Instruct on Hugging Face in July 2026 — reinforces this. Open weights let developers audit the model, fine-tune it for specific use cases, and even run it outside of Windows entirely. It also positions Aion as a legitimate competitor to models like Phi-4 Mini, Mistral 7B, and the Llama family in the broader on-device SLM ecosystem.
What Does This Mean for Regular Windows Users?
You don’t need to be a developer to benefit from Aion 1.0. As it rolls out, you should expect to see it powering:
- Smarter, faster right-click menus in File Explorer with instant summarization
- More capable intent matching in Windows Search
- Better live caption accuracy and real-time accessibility tools
- AI-powered writing suggestions in apps that integrate with Windows AI APIs
- Background agents that can complete multi-step tasks while you do something else
Most of this will be invisible. You’ll just notice that Windows feels smarter, faster at understanding what you mean, and less dependent on a connection.
What Developers Need to Know
If you’re building Windows applications, Aion 1.0 changes the calculus for on-device AI significantly.
Instruct is available now through Edge Insider. You can start building against Windows AI APIs today and expect the feature set to expand when open weights land on Hugging Face in July 2026.
Plan is the longer game. It ships in-box with Windows on capable devices in the coming months. The 32K context window and tool-calling capability mean you can build agentic features that run entirely locally — no API key, no metered usage, no network dependency for the orchestration layer.
The key architectural decision to start thinking about now: which parts of your AI workflow are private or latency-sensitive enough to route locally, and which genuinely need frontier-scale cloud models? Aion makes that hybrid architecture far easier to build on Windows than it has ever been.
Aion 1.0 vs. Apple’s On-Device AI
Apple has built on-device AI deeply into its platform through Apple Intelligence, Foundation Models, and the Neural Engine built into every Apple Silicon chip. It’s been a significant advantage for the Mac and iPhone ecosystem.
Aion 1.0 is Microsoft’s answer — and in one key respect, it has an advantage Apple doesn’t. Aion 1.0 Instruct runs on CPU, which means it works on a far broader range of Windows hardware than Apple’s on-device AI features, which require Apple Silicon. Apple’s on-device AI is better optimized for its specific hardware, but Aion’s reach across billions of existing Windows devices is a different kind of strength.
The comparison will be sharper once both platforms have shipped their full stacks and independent benchmarks are available. For now, what matters is that Windows users finally have a credible on-device AI story — and it’s built into the operating system itself.
Release Timeline
Here’s what’s confirmed and when:
- Now: Aion 1.0 Instruct available in developer preview via Microsoft Edge Insider
- June 9, 2026: Windows Update KB5039239 brings expanded on-device AI stack to Windows 11
- July 2026: Aion 1.0 Instruct open weights release on Hugging Face
- Fall 2026: Aion 1.0 Plan ships in-box with Windows on capable devices, alongside RTX Spark hardware
- Ongoing: Windows AI APIs expanding across more hardware configurations
FAQ: Microsoft Aion 1.0
What is Microsoft Aion 1.0?
Aion 1.0 is a family of small language models built by Microsoft Research and designed to run directly on Windows 11 hardware without cloud connectivity. It was announced at Microsoft Build 2026 on June 2, 2026.
How is Aion 1.0 different from Copilot?
Copilot primarily routes requests to cloud-based models (including OpenAI’s). Aion 1.0 runs entirely on your device — no internet required, no data sent to external servers. They’re designed to complement each other in a hybrid architecture.
What hardware do you need for Aion 1.0?
Aion 1.0 Instruct runs on any modern Windows 11 CPU — no NPU or dedicated GPU required. Aion 1.0 Plan (the 14-billion-parameter model) requires more capable hardware; specific minimum requirements for the Plan model have not yet been officially published.
Is Aion 1.0 open source?
Aion 1.0 Instruct is committed to be released with open weights on Hugging Face in July 2026, allowing developers to fine-tune it and run it outside Windows. Aion 1.0 Plan’s open-source status has not been announced.
What is Aion 1.0 Plan?
Aion 1.0 Plan is a 14-billion-parameter reasoning and tool-calling model with a 32K context window. It enables fully local agentic workflows — AI that can plan multi-step tasks, call Windows APIs, manage files, and orchestrate sub-agents, all without a cloud connection.
When will Aion 1.0 Plan be available?
Microsoft says Aion 1.0 Plan will ship in-box with Windows on capable devices “in the coming months” from the Build 2026 announcement in June 2026. It is expected to arrive alongside RTX Spark hardware in fall 2026.
Does Aion 1.0 replace Phi Silica?
Yes. Aion 1.0 Instruct replaces Phi Silica (Phi-4-mini) as the built-in Windows on-device language model. Microsoft describes it as smaller, faster, and more efficient than its predecessor.
Is my data private with Aion 1.0?
Yes. Because Aion runs locally on your device, your data never leaves your machine during inference. This is one of its primary advantages — you can use AI features on sensitive documents without them being uploaded to a third-party server.
How does Aion 1.0 relate to the Nvidia RTX Spark chip?
Aion 1.0 is the software layer; RTX Spark is the hardware layer. Together they form Microsoft and Nvidia’s platform for agentic AI on Windows. RTX Spark laptops launching in fall 2026 will ship with Aion pre-installed and optimized for the hardware.
Final Thoughts
Microsoft Aion 1.0 is a quiet but genuinely important announcement. It’s the clearest signal yet that Windows is evolving from a platform that runs cloud AI clients into an AI runtime in its own right.
The combination of Aion 1.0 Instruct — which runs on any modern CPU — and Aion 1.0 Plan — a 14-billion-parameter local reasoning model — gives Windows a credible answer to Apple Intelligence for the first time. And the open-weights commitment puts it in competition with the broader SLM ecosystem on Hugging Face.
We’re in the early stages of on-device AI becoming a real part of everyday computing. Aion 1.0 is where that story starts for Windows.
