TL;DR — At Microsoft Build 2026 on June 2, Microsoft AI CEO Mustafa Suleyman introduced 7 new in-house MAI models: MAI-Thinking-1 (reasoning), MAI-Code-1-Flash (coding), MAI-Image-2.5 (image gen + editing), MAI-Voice-2 (TTS), and MAI-Transcribe-1.5 (speech-to-text). All were trained from scratch without distillation from third-party models.

Building a "Hill-Climbing Machine"

Microsoft AI CEO Mustafa Suleyman framed the MAI model family as the first concrete steps toward "Humanist Superintelligence" — AI designed to serve people, not replace them. Critically, every model in the family was trained from scratch on clean, commercially licensed, enterprise-grade data with zero distillation from third-party models including OpenAI or Anthropic. This signals Microsoft is building genuine AI independence.

7 New MAI Models Announced
35B Active Params (MAI-Thinking-1)
43 Languages (MAI-Transcribe-1.5)

The Full MAI Model Lineup

Model Modality Key Feature Availability
MAI-Thinking-1 Text/Reasoning 35B active params MoE, 256K context Private preview
MAI-Code-1-Flash Coding GitHub Copilot + VS Code native GA (individuals)
MAI-Image-2.5 Image Text-to-image + image editing, Arena #3 Foundry now
MAI-Image-2.5 Flash Image Faster/cheaper variant Foundry now
MAI-Voice-2 Speech (TTS) 15+ languages, voice cloning Azure Speech
MAI-Transcribe-1.5 Speech (STT) 43 languages, FLEURS #1 Azure Speech
MAI-Voice-2 Flash Speech Fast variant Coming soon
💡
How does MAI-Thinking-1 compare to the competition?
MAI-Thinking-1 is a sparse Mixture of Experts model with 35B active parameters (~1T total). Microsoft says independent blind raters prefer it over Claude Sonnet 4.6, and it matches Claude Opus 4.6 on SWE-Bench Pro software engineering tasks. It supports a 256K token context window and is built for complex multi-step instructions and code generation.

MAI-Code-1-Flash in GitHub Copilot Now

MAI-Code-1-Flash is a 5B active-parameter lightweight model purpose-built for the GitHub Copilot harness. It rolled out immediately to VS Code Copilot individual users with no setup required. The model uses "adaptive thinking" — staying concise on simple tasks, spending more reasoning budget on complex ones. This directly competes with Anthropic's Claude Haiku at a lower inference cost.

ℹ️
New Hardware: Surface RTX Spark Dev Box
Microsoft also announced the Surface RTX Spark Dev Box, powered by NVIDIA RTX Spark with 128GB unified memory. It can run models up to 120B parameters locally at up to 1 petaflop of AI compute. Targeted at developers needing local agent pipelines and fine-tuning workloads, it will be available later in 2026 in the US.

Microsoft Foundry Becomes an Agent-First OS

Build 2026 turned Microsoft Foundry into a full agent lifecycle platform with three new layers:

Build layer:

  • Toolboxes with MCP-native tool management and auto-auth
  • Work IQ API GA on June 16 — agents can access Microsoft 365 mail, calendar, files

Deploy layer:

  • Hosted Agents in Foundry Agent Service reaching GA in 30 days
  • "Autopilot Agents" with Entra identity, Teams presence, org-chart placement

Operate layer:

  • OpenTelemetry tracing across every model call, tool invocation, and sub-agent hop
  • Agent Optimizer: turns production failures into ranked prompt/skill improvements
Key Takeaways
  • Microsoft publicly signals AI independence with 7 models built without third-party distillation
  • MAI-Thinking-1: 35B MoE reasoning model matching Opus 4.6 on SWE-Bench Pro
  • MAI-Code-1-Flash: live now in GitHub Copilot + VS Code with zero setup
  • MAI-Image-2.5: ranked #3 on Arena.ai for image generation families
  • Microsoft Foundry upgraded to a full agent build/deploy/operate platform
🔗
Resources · Official Sources · Getting Started
— MAI 7 Models Announcement (Microsoft AI)
— MAI-Thinking-1 Deep Dive
— MAI-Code-1-Flash for GitHub Copilot
— MAI Models in Microsoft Foundry — Pricing & Availability