Microsoft launched seven in-house AI models at Build 2026 in San Francisco, every one trained from scratch on commercially licensed data without borrowing outputs from OpenAI, Anthropic, or any other lab, covering reasoning, coding, image generation, voice, and transcription. The flagship reasoning model, MAI-Thinking-1, claims parity with Claude Opus 4.6 on software-engineering benchmarks and a preference edge over Claude Sonnet 4.6 in human evaluations. Every figure came from Microsoft’s own testing, and all the benchmarks await independent replication.
Six of the seven models are in private preview. OpenAI still handles the bulk of production traffic across Microsoft 365 Copilot and GitHub Copilot.
Seven MAI Models, One Hill-Climbing Pipeline
The family covers five functional areas, announced in a single keynote by Mustafa Suleyman, who runs Microsoft AI:
- MAI-Thinking-1: a sparse Mixture of Experts (MoE, a neural-network design where only a fraction of parameters activate per request to control inference cost) architecture, with 35 billion active parameters out of approximately 1 trillion total. Context window of 256,000 tokens, enough, Microsoft says, to process a 600-page document in one pass. In private preview through Microsoft Foundry; pricing not announced.
- MAI-Code-1-Flash: 5 billion parameters, tuned on GitHub Copilot’s production tool harnesses rather than academic benchmarks. Rolling out gradually to all Free, Pro, Pro+ and Max GitHub Copilot plans; already the default model in VS Code.
- MAI-Image-2.5 and a flash variant: text-to-image and image-to-image generation, active in PowerPoint and rolling into OneDrive.
- MAI-Voice-2 and a flash variant: text-to-speech across more than 15 languages with voice cloning. MAI-Transcribe-1.5 handles speech recognition in 43 languages, already live in Azure Speech.
All seven are distributed through Azure AI Foundry and, notably, on Fireworks AI, Baseten, and Open Router, departing from Azure-only delivery. Microsoft co-designed the models with its Maia 200 inference accelerator, reporting a 30% performance-per-dollar improvement against Nvidia’s GB200 and a 1.4x performance-per-watt gain running the MAI family on Maia end-to-end. Both figures are Microsoft’s, pending external verification.
Suleyman described the training program as a
