In a nutshell: Microsoft has introduced MAI-Thinking-1, its first reasoning model with fine-tuning capability for enterprise, specifically designed for domain-specific customizations.
Microsoft unveiled a new model family with seven MAI models at the Build conference, led by MAI-Thinking-1 as a reasoning model. Technical details show a 35-billion-parameter mixture-of-experts model with a 256K-token context window, trained on 30 trillion tokens.
Microsoft presented the MAI model family with seven new models at the Build conference, covering reasoning, code generation, image editing, speech transcription, and voice synthesis. The flagship models are MAI-Thinking-1 (reasoning), MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, and MAI-Voice-2. The company positioned itself as both an AI platform provider and a frontier model laboratory.
MAI-Thinking-1 is described as Microsoft’s first reasoning model, trained with clean data provenance without distillation from third-party models. The model is a mixture-of-experts architecture with 35 billion active parameters, a 256K-token context window, and was trained on 30 trillion tokens using 8,192 GPU200 GPUs. Microsoft reports performance of 97 percent on the AIME-2025 benchmark and 53 percent on the SWE-Bench-Pro benchmark, with Surge testers preferring it over Claude Sonnet 4.6 in direct comparison.
For CIOs and CTOs, MAI is significant because these models are specifically designed for domain-specific fine-tuning — a clear distinction from frontier labs that have largely phased out fine-tuning. This underscores Microsoft’s strategy of enabling enterprise customers with localized and customized AI capabilities. With the acquisition of Inflection two years ago, Microsoft has built a “Tier-2 neolab” that is relevant for specialized enterprise workloads.
Microsoft additionally released a 109-page technical report on MAI-Thinking-1, which was recognized by the research community for its transparency level. The Build announcements also included Windows AI features, a new Surface RTX Spark dev box, GitHub Copilot improvements with Canvas and cross-device continuity support, and Web-IQ as a new grounding and search API stack for AI agents.
Source: www.latent.space · Published June 3, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.2.9.