Microsoft unveils Aion 1.0 on-device AI models at Build 2026

Microsoft Unveils Aion 1.0 AI Models for Windows 11 at Build 2026

Microsoft used the opening keynote of its Build 2026 developer conference on Tuesday to announce Aion 1.0, a new family of on-device AI models designed to run locally on Windows PCs, marking an expansion of the company’s push to embed intelligence directly into the operating system without relying on cloud services.

New Models for Local AI

CEO Satya Nadella introduced two models in the Aion 1.0 lineup. Aion 1.0 Instruct is a small language model that replaces the previous Phi-4-mini in Microsoft Edge, offering a smaller and faster footprint that expands compatibility to devices with less capable GPUs and even those relying solely on CPU inference. A developer preview is now available in Edge Canary and Dev channels, with a planned open-source release on Hugging Face in July.windows

Aion 1.0 Plan is a 14-billion-parameter reasoning and tool-calling model aimed at agentic workflows. Built into Windows, it enables applications to reason over user intent, invoke tools, manage files, and coordinate sub-agents directly on local hardware. Microsoft said the model will arrive in the coming months.firstpost

Expanding Windows AI Across Hardware

Alongside the models, Microsoft announced an expansion of its Windows AI APIs to run across CPUs, GPUs, and NPUs from AMD, Intel, Qualcomm, and Nvidia. The move allows developers to tap local compute rather than defaulting to cloud-based inference, improving privacy, reducing latency, and eliminating per-query costs.stocktwits

Tom’s Guide reported that Nadella framed the initiative as taking “AI use and development away from the cloud and to one’s PC,” with processors from all four major silicon vendors enabling the transition.tomsguide

Hardware and Developer Tools

The conference also introduced the Surface RTX Spark Dev Box, a compact developer workstation powered by Nvidia’s RTX Spark silicon. The device delivers up to 1 petaflop of AI performance with 128 gigabytes of unified memory, capable of running models with up to 120 billion parameters locally within a 100-watt power envelope.firstpost

Build 2026 runs June 2–3 at Fort Mason Center in San Francisco, with sessions spanning agentic AI, Azure AI Foundry, GitHub Copilot enhancements, and responsible AI tooling.notebookcheck