Microsoft Tests Homegrown AI Model MAI-1-preview to Challenge OpenAI and Boost Copilot

Microsoft has begun publicly testing a new in-house AI model that could mark a pivotal shift in its relationship with OpenAI and intensify competition in the AI space. The company’s MAI-1-preview model is now available for evaluation on LMArena, a platform where users can compare AI performance across different models. Microsoft plans to gradually roll out MAI-1-preview for select text-based tasks within its Copilot assistant over the coming weeks, using real-world user feedback to refine the model. The company has also launched a form for developers to request early access to the model, signaling its intent to build a community around its growing AI capabilities. While Microsoft remains a major investor in OpenAI—having poured over $13 billion into the startup and providing the cloud infrastructure that powers many of its models—it is now advancing its own AI ambitions. The move comes as the two companies increasingly operate in parallel, with Microsoft now listing OpenAI among its competitors in its annual report, a notable shift from earlier years when the partnership was seen as a singular strategic alliance. OpenAI has also diversified its cloud partnerships in recent months, turning to providers like CoreWeave, Google, and Oracle to meet rising demand for its models. Its ChatGPT assistant now reaches 700 million users weekly, underscoring the scale of its success. On LMArena, MAI-1-preview ranked 13th for text workloads as of Thursday, falling behind models from Anthropic, DeepSeek, Google, Mistral, OpenAI, and xAI. Still, Microsoft emphasized that this is an early version of a model trained end to end in-house, a milestone for its AI development efforts. The model was developed using approximately 15,000 Nvidia H100 GPUs and includes a dedicated cluster of the newer Nvidia GB200 chips, highlighting Microsoft’s growing investment in high-performance AI compute infrastructure. Mustafa Suleyman, CEO of Microsoft’s AI division, announced the launch on X, stating that MAI-1-preview represents “our first foundation model trained end to end in house.” Suleyman, a former co-founder of DeepMind and a key figure in the AI world, joined Microsoft in 2023 after leading Inflection AI, which was acquired by Microsoft. His team has since expanded rapidly, bringing in around two dozen former DeepMind researchers. Microsoft has also been developing smaller, open-source language models under the Phi series, but MAI-1-preview marks a significant leap toward building a full-scale, proprietary foundation model. Suleyman described the project as part of a broader vision to advance AI models, scale compute capabilities, and deliver transformative experiences to billions through Microsoft’s products.