Microsoft used its Build 2026 developer conference to show off a new family of internal AI models, signaling a bigger move toward building and running more of its own artificial intelligence systems. The company said the new MAI lineup includes seven models developed in-house for tasks ranging from reasoning and coding to image generation, transcription and speech synthesis.
According to Microsoft AI chief Mustafa Suleyman, the company’s ability to train its own frontier models improved after contract terms with OpenAI were renegotiated about six months ago. He said the revised arrangement removed a restriction that had limited Microsoft from training models at the cutting edge on its own.
The flagship release is MAI-Thinking-1, a 35-billion-parameter reasoning model trained from the ground up using licensed and cleaned data. Microsoft says it did not rely on distillation from another lab’s model for that system, a detail that sets it apart from some other AI efforts in the market.
The rest of the MAI family is aimed at specific product needs. MAI-Code-1-Flash is intended for GitHub Copilot. MAI-Image-2.5 handles text-to-image generation and editing. MAI-Transcribe-1.5 supports speech recognition across 43 languages, while MAI-Voice-2 is built for multilingual voice output. Microsoft said the models are available through Microsoft Foundry and can be adapted by companies for their own workflows.
To help businesses customize them, Microsoft is promoting a capability it calls Frontier Tuning. The company describes it as a reinforcement learning environment that lets enterprises tailor models to proprietary tasks.
Beyond model releases, Microsoft also framed the announcement as part of a larger shift toward agentic AI. Suleyman introduced the idea of an “Actions Quotient,” or AQ, as a measure of how well systems can carry out tasks, not just answer questions.
Microsoft Scout was presented as the company’s first Autopilot agent. The tool is designed to run continuously as an assistant and uses governed identity features in Microsoft Entra. Microsoft also said its Foundry platform now supports agents with very fast startup times and one-click deployment into products such as Teams and Copilot.
The company pointed to an example of Frontier Tuning in action, saying a model adapted for Excel matched the performance of GPT 5.4 while costing about one-tenth as much. Microsoft did not provide additional technical details in the source material, but the comparison suggests the company is positioning customization as a way to lower AI operating costs for enterprise customers.
Microsoft also connected the MAI rollout to its hardware strategy. The company is one of the largest buyers of GPUs in the world, but it is also developing its own chips. It said the Maia 200 accelerator is 30% more cost-efficient than Nvidia’s GB200, and that running optimized MAI models on the chip improves performance per watt by another 1.4 times.
Taken together, the announcements show Microsoft pushing more aggressively to control the full AI stack, from models to agents to hardware. That strategy could reduce reliance on external partners and give the company more flexibility as competition intensifies in enterprise AI.