MAI

MAI is the name Microsoft AI gives to the models it trains in-house, as distinct from the OpenAI models that Microsoft products have historically relied on. The family began on 28 August 2025 with two releases: MAI-Voice-1, a speech generation model used in Copilot Daily and Podcasts, and MAI-1-preview, a mixture-of-experts text model described as the division’s first foundation model trained end to end, which Microsoft began testing publicly and routing some Copilot queries to. MAI-Image-1, a text-to-image model, followed in October 2025.

In June 2026 Microsoft AI released seven models at once under the banner of building “a hillclimbing machine”: MAI-Thinking-1, a medium-sized flagship reasoning model; MAI-Code-1-Flash, an inference-efficient coding model built into GitHub Copilot and VS Code; MAI-Image-2.5 and a Flash variant; MAI-Transcribe-1.5 for speech-to-text in dozens of languages; and MAI-Voice-2 with a lower-cost Flash variant. Microsoft says these models are trained without distilling from other labs’ models and are co-designed with its Maia silicon. The company’s models page now lists updated point releases across the same lines, plus MAI-Cyber-1-Flash, a cybersecurity model used in its vulnerability-remediation harness.

The family spans text reasoning, code, image generation and editing, speech synthesis and transcription. The models are offered through Microsoft Foundry, through third-party providers including OpenRouter, Fireworks and Baseten, and are built into Microsoft products. Microsoft has said developers can tune some model weights, but the models are not released as open weights.

MAI is the model program of the division that houses the MAI Superintelligence Team, formed in November 2025 under Mustafa Suleyman to pursue what he calls “humanist superintelligence.” Microsoft has not claimed that any MAI model is one.

For current details see microsoft.ai/models.