At the recently opened Build2026 Developer Conference, Microsoft unveiled a series of self-developed AI models. The launch of its first advanced reasoning model, MAI-Thinking-1, marks a crucial step for Microsoft in independent large model development and building a comprehensive AI matrix across all scenarios.

As a "medium-sized model" with 35 billion active parameters, MAI-Thinking-1 has reached industry-leading levels in key software engineering benchmark tests. Microsoft emphasized that the model was trained from scratch using clean data, and explicitly committed to not using any distilled data from third-party models. In an industry where data distillation is widely used, this demonstrates Microsoft's core commitment to technical autonomy and data purity.

In addition to the major reasoning model, Microsoft also expanded the MAI family lineup for multimodal and vertical application scenarios. In the fields of image and voice, MAI-Image2.5, which supports text-to-image generation and image editing, along with its Flash version, was officially released; MAI-Transcribe-1.5, which features ultra-fast transcription capabilities, can reach five times the speed of competitors; and MAI-Voice-2, which now supports 15 additional languages (including an upcoming Flash version), significantly enriches the voice multimodal ecosystem.

For the developer ecosystem, the coding model MAI-Code-1, which features optimized reasoning efficiency, has been successfully integrated into GitHub Copilot and Visual Studio Code. This successful rollout of multiple self-developed models not only completes Microsoft's closed-loop ecosystem from underlying reasoning to upper-level applications but also highlights the strategic determination of a tech giant to pursue comprehensive AI self-reliance and reduce reliance on external technologies.