At Build 2026, Microsoft’s AI Superintelligence Team introduced seven in-house MAI models. The centerpiece, MAI-Thinking-1, is a reasoning model with 35 billion active parameters and a 256,000-token context window.
Ce qui change
The rest of the lineup covers image generation, speech synthesis, transcription and coding, including MAI-Image-2.5, MAI-Voice-2, MAI-Transcribe-1.5, MAI-Code-1 and MAI-Code-1-Flash.
À surveiller
Microsoft says a version tuned for McKinsey outperformed OpenAI’s GPT-5.5 with ten times better cost efficiency. The message to developers and to OpenAI is clear: Microsoft wants its own cheap model stack, not only a downstream license.
Sources：CNBC, Yahoo Finance and Windows Central reports on Microsoft MAI models；CocoLoop