原文 · 未翻译
Build 2026: Microsoft tops Google in image generation while playing catch-up on reasoning
Key Points
Microsoft unveiled seven homegrown AI models at Build 2026, including its first reasoning model, MAI-Thinking-1. In benchmarks, it lands roughly on par with Deepseek V3.2.
A new method called "Frontier Tuning" lets companies adapt models to their own workflows using reinforcement learning. Microsoft says tuned models match GPT-5.4 performance at one-tenth the cost.
Microsoft is also launching "Scout," an always-on background agent that handles office tasks like scheduling and meeting prep. The software announcements are paired with local developer hardware and a new operating system built for AI agents.
At Build 2026, Microsoft announced seven new AI models developed in-house, including its first reasoning model. The company also introduced a new tuning method and an autonomous background agent.
The centerpiece is MAI-Thinking-1, Microsoft's first reasoning model. According to Microsoft AI chief Mustafa Suleyman, it's a 1-trillion-parameter model with 35 billion active parameters and a 128,000-token context window, built for multi-step instructions, long contexts, and code generation.
Microsoft says MAI-Thinking-1 matches leading models on key software engineering benchmarks and was preferred over Anthropic's Sonnet 4.6 in internal blind comparisons. The model was trained from scratch on clean data without distillation from third-party models, according to Suleyman. That's a not-so-subtle jab at practices at other labs. A look at the published benchmarks, though, puts the model roughly on par with Deepseek V3.2.
A model family spanning six task areas
Beyond the reasoning model, the MAI family includes six more systems. MAI-Code-1-Flash is an agentic coding model with 5 billion parameters that Microsoft says is comparable to Anthropic's Haiku but cheaper to run. It's integrated into GitHub Copilot and Visual Studio Code.
MAI-Image-2.5 handles text-to-image and image editing, landing second place on the Arena-Score image benchmark behind GPT-Image-2 and ahead of Google's Nano-Banana models. MAI-Transcribe-1.5 is pitched as the fastest transcription model, supporting 43 languages. MAI-Voice-2 generates speech in 15 languages and can clone voices from short samples.
All models share the same data foundation, infrastructure, and evaluation pipeline, according to Microsoft. They're available through Azure Foundry, and for the first time, developers can fine-tune the weights themselves.
Frontier Tuning makes the cost case
Microsoft is pairing the models with a new approach called Frontier Tuning. Customers can use reinforcement learning environments to align models directly with their own workflows. The most valuable training data, Microsoft argues, is the actual work traces an agent leaves behind inside an organization.
In an internal test, a MAI model tuned for Excel matched GPT-5.4's performance while running up to ten times more efficiently. At McKinsey, a customized MAI model achieved the highest win rate of any system tested, again at roughly one-tenth the cost.
Scout is Microsoft's first always-on agent
The third pillar is a new agent category Microsoft calls "Autopilots." These are persistent agents with their own identity that work autonomously in the background. The first one is Microsoft Scout, integrated into Teams, Outlook, OneDrive, and SharePoint.
Scout is designed to coordinate meetings across time zones, prepare briefing materials, schedule upcoming deliverables on your calendar, and flag stalled decisions before they become blockers. Through a component called Work IQ, the agent builds a context memory of how you work and what you prioritize.
Each agent runs under its own Entra identity with tightly scoped access rights, sandboxed execution via Microsoft Execution Containers, and mandatory human approval for sensitive actions. Credentials are also scoped to each task and scrubbed from logs. Whether that's enough remains to be seen. Previous agent systems have consistently failed at exactly the point where language models meet external data.
Scout is available first as an experimental release through the Frontier program. It requires an Intune configuration and a GitHub Copilot license.
Hardware, an OS, and a clinical model round out the strategy
The software announcements come alongside several more pieces of a broader AI strategy. With Project Solara, Microsoft is previewing an Android-based operating system designed to run agents across devices, co-developed with Qualcomm and MediaTek. On the Build stage, the company showed a desktop hub and a digital badge as possible form factors.
For local AI development, Microsoft is launching the Surface RTX Spark Dev Box, equipped with Nvidia's Arm-based Spark RTX chip and 128 GB of unified memory. Pricing and full specs haven't been announced yet.
In healthcare, Microsoft announced a partnership with the Mayo Clinic to co-develop a clinical foundation model. The model will first be deployed in Mayo Clinic's own operations and later made available through Azure Foundry. The Mayo Clinic retains ownership.
Microsoft frames the overarching goal as "Humanist Superintelligence," meaning AI systems that remain tools under human control. Suleyman says the company plans to rapidly scale compute and capabilities over the coming year, backed in part by Microsoft's own Maia 200 chips.
AI News Without the Hype – Curated by Humans