Microsoft Builds Its Own MAI Models to Reduce OpenAI Reliance
Across the first half of 2026 Microsoft shipped its own MAI model family, from Foundry-available speech and image models in April to seven new models led by MAI-Thinking-1 at Build in June, plus the open-weight Phi-4-reasoning-vision.
Through the first half of 2026 Microsoft moved from being primarily a consumer of OpenAI’s models to shipping a growing family of its own, called MAI. At Build on 2 June 2026 it unveiled seven new in-house MAI models led by MAI-Thinking-1, its first in-house reasoning model, after opening its first MAI models to developers in April. Microsoft states the models were trained from scratch without distillation from third-party models, which is the strategic point: reducing dependence on OpenAI.
What happened
The rollout came in stages. On 2 April 2026 Microsoft released its first in-house models into Microsoft Foundry for third-party developers: MAI-Transcribe-1 (multilingual speech-to-text), MAI-Voice-1 (fast expressive text-to-speech with voice cloning), and MAI-Image-2 (its highest-capability text-to-image model), the same models already powering Copilot and Bing. On 4 March 2026 Microsoft Research had released Phi-4-reasoning-vision, a 15-billion-parameter open-weight multimodal reasoning model that extends the small-model Phi line into vision and document understanding.
At Build on 2 June 2026, Microsoft AI unveiled the broader MAI family, led by MAI-Thinking-1 (a mid-weight reasoning model) alongside image, speech, voice, and code models. The consistent message is that these are Microsoft’s own models, trained in-house, positioned to run its products and to be offered to developers through Foundry.
Why it matters for builders
For years the Microsoft AI story was Azure plus OpenAI. A full in-house stack, spanning reasoning, image, speech, and code, changes the relationship: Microsoft now has models it controls end to end, which affects pricing, availability, and roadmap independence for anyone building on its platforms. It also adds another serious reasoning model option to evaluate.
The open-weight Phi-4-reasoning-vision matters for a different reason: it is small enough to self-host and run on modest hardware, which suits agentic and document workflows where you want control and low cost rather than a frontier API. If you build on Microsoft Foundry, you now have first-party MAI models alongside OpenAI and Anthropic models in one place; benchmark them against your workload rather than assuming the default. See the 2026 LLM landscape for the wider field.
Sources
- Microsoft AI, “Building a hill-climbing machine: launching seven new MAI models” (2 June 2026): https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/
- Microsoft AI, “Announcing 3 new world-class MAI models available in Foundry” (2 April 2026): https://microsoft.ai/news/today-were-announcing-3-new-world-class-mai-models-available-in-foundry/
- Microsoft Research, “Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model” (4 March 2026): https://www.microsoft.com/en-us/research/blog/phi-4-reasoning-vision-and-the-lessons-of-training-a-multimodal-reasoning-model/
Further reading
- The 2026 LLM landscape : where MAI sits among the frontier and open models.
- What are reasoning models? : the class MAI-Thinking-1 belongs to.
- Claude Opus 4.8 : one of the third-party models still offered alongside MAI in Foundry.