Content
summary Summary

Microsoft is reportedly developing its own family of language models under "CEO of AI" Mustafa Suleyman, aiming to match capabilities offered by OpenAI and Anthropic. The company is said to be planning to release these models through an API later this year.

Ad

The Information reports that Microsoft's AI team reached a key milestone with their internal model family, known as MAI. Testing indicates the models perform at nearly the same level as leading offerings from OpenAI and Anthropic on standard benchmarks. The family includes a reasoning model specifically designed to match OpenAI's o1 capabilities.

These new models are much larger than Microsoft's earlier Phi series, which focused more on balancing cost and performance. Microsoft could open up API access to outside developers before year's end, putting it in direct competition with OpenAI and other AI labs offering similar services.

Microsoft isn't limiting itself to internal development - the company is already evaluating models from xAI, Meta, and DeepSeek in Copilot as potential replacements for its current OpenAI models. This strategy could help reduce Microsoft's reliance on its AI partner.

Ad
Ad

Building independent AI capabilities

The path to developing the MAI models hasn't been straightforward, The Information reports. Over its year-long development, the project encountered technical difficulties, changes in direction, and the departure of several key team members, all while OpenAI continued releasing new versions of its models.

The departure of Sébastien Bubeck, who headed the Phi project, hit especially hard when he left for OpenAI, taking several Microsoft researchers with him.

Suleyman's frustration with OpenAI grew during this period, particularly because OpenAI wouldn't reveal details about their o1 model's inner workings. Still, the AI team under Karén Simonyan managed to develop comparable reasoning abilities using chain-of-thought techniques.

Reports from spring 2024 mentioned a Microsoft model called MAI-1 with approximately 500 billion parameters. While Microsoft's CTO Kevin Scott confirmed MAI's existence on LinkedIn, he stressed the company's continuing partnership with OpenAI. The new model incorporates technology and training data from Suleyman's AI startup Inflection, which Microsoft acquired for $650 million.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Microsoft is developing its own AI model family called MAI under the leadership of Mustafa Suleyman, aiming to achieve performance comparable to leading models from OpenAI and Anthropic in benchmarks.
  • The MAI models are significantly larger than Microsoft's previous Phi family and could be made available as an API for external developers within the year, with initial tests already underway to integrate MAI into Copilot, replacing the current OpenAI models.
  • The development process for the MAI models spanned nearly a year, marked by technical setbacks, strategic changes in direction, and the departure of key talent to OpenAI.
Sources
Matthias is the co-founder and publisher of THE DECODER, exploring how AI is fundamentally changing the relationship between humans and computers.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.