MAI family
Also known as: Microsoft AI models, MAI models, MAI-Thinking-1, MAI-Code-1-Flash, MAI model family
Launched June 2, 2026 at Microsoft Build, the MAI family is Microsoft AI's first meaningful move from AI distributor to AI model maker. All seven models were trained from scratch on clean, commercially licensed data with no distillation from third-party models, a deliberate choice Microsoft frames as both an IP risk mitigation and a quality control stance. The family covers five modalities: reasoning, coding, image generation and editing, voice synthesis, and speech transcription.
MAI-Thinking-1, the headline model, is a mixture-of-experts reasoning model with 35 billion active parameters and a 256K context window, sitting in a mid-sized weight class and targeting complex multi-step instructions, long-context reasoning, and code. MAI-Code-1-Flash is tuned specifically for GitHub Copilot and Visual Studio Code workflows. MAI-Image-2.5 supports both text-to-image and image-to-image tasks. The family is co-designed with Microsoft's own Maia 200 silicon, which provides a reported 1.4x performance-per-watt improvement over third-party hardware.
Strategically, the MAI launch signals that Microsoft is no longer content to be purely an OpenAI distribution channel. All MAI models are available through Microsoft Foundry and also on third-party inference providers including OpenRouter, Fireworks AI, and Baseten. The most consequential pairing is MAI models with Frontier Tuning, Microsoft's enterprise RL system, which allows organizations to adapt MAI models to their own workflows while keeping the learning loop inside their compliance boundary.