Microsoft launches 7 self-developed AI models, reducing OpenAI reliance

Microsoft is making a significant push into the artificial intelligence market with the launch of seven self-developed AI models. Unveiled at the annual Build developer conference in San Francisco, these models aim to reduce the company's dependence on OpenAI's technology.

The centerpiece is the 'MAI-Thinking-1' model, a mid-sized model with 35 billion parameters. Microsoft emphasizes that this model goes beyond simple knowledge recall, possessing the ability to verify its own answers and solve problems logically. The company highlighted that the model was trained entirely on its own data, without using "distillation" from other AI models' results. In blind tests, MAI-Thinking-1 reportedly surpassed Anthropic's Claude 4.6 Sonnet and demonstrated superior efficiency compared to OpenAI's GPT-5.5, using up to 10 times less token usage for data processing.

“

MAI-Thinking-1 is a mid-sized model with 35 billion parameters that goes beyond simply memorizing knowledge to answer questions, and can verify its own answers and solve problems logically.

Microsoft also introduced specialized models for business applications. These include 'MAI-Code-1-Flash,' a coding model for developers, and 'MAI-Voice-2,' designed to generate speech with natural emotional expression. The image generation model, 'MAI-Image-2.5,' was touted as outperforming Google's 'Nano Banana Pro' in performance evaluations.

This move is seen as a strategic effort to lessen Microsoft's reliance on OpenAI, with whom it has had a close partnership since 2019, investing over $13 billion by 2023. While Microsoft has leveraged OpenAI models for its Copilot AI assistant, the recent announcement of the end of their exclusive partnership foreshadowed this competitive pivot. The introduction of its own robust AI models positions Microsoft to compete more directly in the rapidly evolving AI landscape.

“

In blind tests, it surpassed Anthropic's Claude 4.6 Sonnet and reduced token usage by up to one-tenth compared to OpenAI's GPT-5.5 model during data processing, demonstrating high efficiency.