Microsoft launches its own AI model MAI-Thinking-1, expanding reasoning and multimodal capabilities

At the Build 2026 developer conference, Microsoft unveiled its in-house AI model MAI-Thinking-1, a medium-sized reasoning model with 35 billion active parameters. Microsoft says it was trained from scratch on clean data without using any third-party distillation, and it matches top models on software engineering benchmarks.

Alongside the reasoning model, Microsoft launched several other models for different tasks: MAI-Image 2.5 for text-to-image and image editing; MAI-Transcribe-1.5, which is up to five times faster than competitors for speech transcription; MAI-Voice-2 with support for 15 new languages; and MAI-Code-1, already integrated into GitHub Copilot and Visual Studio Code with optimized inference efficiency.

This release marks a major step for Microsoft in expanding its in-house MAI model family across reasoning, image, voice, and coding AI applications.