Particle.news
Download on the App Store
80 ARTICLES
·
last wk.

Microsoft Announces End-to-End AI Stack With New Models, Agents and Hardware

The company says the push will cut dependence on external models, lower cloud metering costs, and put agentic AI under enterprise control.

Image
Image
Image
Project Solara devices on a virtual shelf

Overview

  • At Build 2026 in San Francisco on Tuesday, Microsoft unveiled a coordinated suite of AI products that span proprietary models, always-on agents, developer hardware, device software, sandboxing and quantum research.
  • The company introduced seven MAI models led by MAI-Thinking-1, a 35-billion-parameter reasoning model with a 128K context window built for long-context reasoning and code generation.
  • Microsoft showed Scout, an always-on Autopilot agent built on the OpenClaw framework that will run in a limited desktop preview for Frontier customers and is slated to be the first in a Copilot super app roadmap.
  • New hardware and on-device work include the Surface RTX Spark Dev Box with Nvidia’s Spark RTX silicon, Aion and Phi Silica local models for planning and background tasks, and Project Solara prototypes for agent-first devices intended to enable 'unmetered' local AI.
  • The company also announced Microsoft Execution Containers and Entra identities to sandbox agents and give IT observable, enforceable controls, while most products remain in private or limited preview and wider availability and pricing were not disclosed.