At Build 2026, Microsoft announced Aion 1.0 - two small language models baked directly into Windows 11. Aion 1.0 Plan is a 14-billion-parameter reasoning and tool-calling model with a 32K context window that ships in-box on capable devices. That last phrase - "capable devices" - is doing a lot of quiet work. The NPU path requires at least a 40-TOPS NPU, which at launch means Qualcomm Snapdragon X Elite or Intel Lunar Lake. Intel Meteor Lake's NPU (~10-11 TOPS) does not meet the floor. If your team is still on Meteor Lake machines - common in fleets refreshed in late 2023 and 2024 - Aion 1.0 Plan is not available to them yet, regardless of what the Build keynote implied.
That is the gap most coverage skips. Aion is a real, interesting release. It is also gated behind a hardware refresh cycle that most enterprise IT departments have not completed.
What Aion 1.0 Instruct and Plan actually do
Aion 1.0 is two distinct models with different jobs, not a single release.
Aion 1.0 Instruct is Microsoft's next-generation small language model, described as smaller, faster, and more efficient than the previous Windows OS SLM. It is designed for on-device workloads: summarization, rewriting, intent detection, and accessibility features.
Microsoft has not published a parameter count for Aion 1.0 Instruct , which matters if you want to compare it against Phi-4-mini or Gemma 4 E4B on identical tasks. Developers can experiment with Aion 1.0 Instruct in preview in Edge Insider channels, with open weights on Hugging Face committed for July. Those weights are now past their scheduled date.
Aion 1.0 Plan is a 14-billion-parameter reasoning and tool-calling model with a 32K context window. It is built to reason over user intent, call tools, manage files, and orchestrate sub-agents - a fully local agent loop.
Microsoft says it is coming "in the coming months," so it is announced, not yet shipped.
Both models run on top of Windows ML, the on-device inference runtime that went generally available in 2025 and now underpins Windows AI Foundry and Foundry Local. Under the hood, that is DirectML plus the ONNX Runtime spreading work across CPU, GPU, and NPU.
The hardware gate that most teams will hit
The NPU path requires a Copilot+ PC with at least a 40-TOPS NPU. At launch that means Qualcomm Snapdragon X Elite (45 TOPS) or Intel Lunar Lake (45-48 TOPS). AMD's Ryzen AI chips (50 TOPS) qualify on hardware but Aion NPU support is deferred. Intel Meteor Lake's NPU (~10-11 TOPS) does not meet the floor.
That is a significant exclusion. Meteor Lake shipped broadly across commercial laptops through 2024, meaning a large slice of managed Windows fleets bought in the last hardware cycle cannot run Aion 1.0 Plan on its NPU path at all. The models can also run on capable discrete GPUs via Windows ML
- but most business laptops do not have a discrete GPU worth running a 14B model on.
Teams evaluating deployment across managed Windows fleets should flag hardware specification disclosure as a prerequisite for planning. Specifically: Microsoft has not yet disclosed minimum RAM or GPU specs for Aion 1.0 Plan. Enterprise IT teams cannot scope a rollout without that data.
What is genuinely new versus incremental
The honest answer is: the delivery model is new; the model capability claims are plausible but unverified.
For most of its history, Windows has been a platform for running other people's software. At Build 2026, Microsoft made its most direct move to change that: it is now shipping its own AI models as first-class components of the operating system itself, bundled with Windows the way a font renderer or filesystem driver is bundled.
Shipping Aion 1.0 Instruct with open weights is a strategic choice as much as a technical one. Open weights let developers fine-tune, audit, and embed the model without a licensing negotiation. For a company that spent years tightly coupled to a single frontier partner, opening up a Windows-native model is a notable shift toward owning the stack on its own terms.
What is incremental: the underlying tasks - summarization, rewriting, intent detection - are the same jobs Phi-4-mini and Gemma 4 E4B already handle well. Phi-4-mini, for instance, is 3.8B parameters, downloads as a 2.49 GB GGUF at Q4_K_M, handles a 128,000-token context window, and scores 88.6% on GSM8K math benchmarks
- all available today via Ollama without needing a specific chip generation. The Aion advantage is integration, not raw capability: it runs through Windows AI APIs rather than a separately installed runtime, which matters for apps that need to call it from JavaScript or a browser extension.
Edge 148's on-device Translator API, separate from Aion but part of the same Build announcement, is available now - running across 145+ languages, hardware-accelerated on NPUs and CPUs, with zero translation costs compared to cloud-based services. That is the part of the announcement that is actually shipped and available today regardless of your laptop generation.
What to actually do with this information
If you are evaluating private, in-house AI for your team right now:
- Check your hardware first. The NPU path requires a Copilot+ PC with at least a 40-TOPS NPU - that means Qualcomm Snapdragon X Elite (45 TOPS) or Intel Lunar Lake (45-48 TOPS) at launch. Pull your fleet's processor list before scoping a rollout.
- Aion 1.0 Instruct is testable today. You can test it through the Windows Copilot Runtime API if you're on an Edge Insider build. The open weights are past their July commitment date - check Hugging Face for the microsoft/aion-1.0-instruct repository directly.
- Aion 1.0 Plan is not yet available. Full agentic workflows running locally on Windows with no cloud dependency are the goal, but Plan has no confirmed GA date. Do not build against it yet.
- Hold the comparison to a real baseline. Before committing to the Windows-native path, run the same task through Phi-4-mini via Ollama on the same machine. The integration story is real; the quality advantage is still unproven.
A teammate like Beagle - running inside Slack or Teams - can surface answers from your existing docs without touching any external server, which covers a lot of the same privacy ground for the workplace communication layer specifically. See how AI agents in Slack actually handle your data if that is the real concern.
Beagle in action#it-ops, 10:22amAion 1.0 on-device AI: common questions
What is Microsoft Aion 1.0?
Aion 1.0 is a family of on-device small language models for Windows 11, announced at Build 2026 on June 2. It has two members: Aion 1.0 Instruct, a small SLM for summarization and rewriting currently in preview, and Aion 1.0 Plan, a 14B reasoning and tool-calling model with a 32K context window coming in the following months.
What hardware do I need to run Aion 1.0?
The NPU path requires a Copilot+ PC with at least a 40-TOPS NPU. At launch that means Qualcomm Snapdragon X Elite (45 TOPS) or Intel Lunar Lake (45-48 TOPS). AMD's Ryzen AI chips (50 TOPS) qualify on hardware but NPU support is deferred. Intel Meteor Lake (~10-11 TOPS) does not meet the floor.
How is Aion 1.0 different from Phi-4-mini?
Aion 1.0 is Microsoft's OS-native path: callable via Windows AI APIs and Edge, no separate runtime needed. Phi-4-mini is an open-weight model you run yourself, typically via Ollama or llama.cpp. Phi-4-mini handles a 128,000-token context window at Q4_K_M quantization and is available on any machine today. Aion 1.0 Instruct's parameter count is undisclosed, so a head-to-head quality comparison is not yet possible.
Are Aion 1.0 weights open source?
Aion 1.0 Instruct ships with open weights. Microsoft committed to releasing it as open source on Hugging Face in July 2026. Aion 1.0 Plan's weights have not been announced as open.
Is Aion 1.0 private - does my data leave the device?
Aion 1.0 includes Aion 1.0 Instruct and Aion 1.0 Plan, both running locally so data does not have to leave the device. That is the design intent. Read what a specific feature actually does with your data: a local inference step does not automatically protect everything downstream of it. Verify that any app calling Aion 1.0 is not syncing outputs to a cloud endpoint afterward.