Microsoft has started public testing of its first in-house foundation model, MAI-1-preview, marking a significant step in reducing reliance on OpenAI for powering AI features across its products. The model is being evaluated on LMArena, where users can rank AI systems on performance across text-based tasks.
The company said MAI-1-preview will be gradually integrated into Copilot for select text use cases in the coming weeks, with feedback helping refine future iterations. Early benchmarks placed the model 13th on LMArena, trailing offerings from Anthropic, DeepSeek, Google, Mistral, OpenAI, and xAI.
MAI-1 was trained with the support of 15,000 Nvidia H100 GPUs, alongside a working cluster of Nvidia GB200 chips. Microsoft AI CEO Mustafa Suleyman, who previously co-founded DeepMind and led Inflection AI, described the model as the company’s first large-scale end-to-end system developed entirely in-house.
While Microsoft remains a key investor in OpenAI, with over $13 billion committed, the launch of MAI-1 signals a shift as both firms increasingly compete in the fast-evolving AI landscape.
