Microsoft is advancing open-weight AI models as part of a broader strategy to enhance its Azure cloud platform. The company, along with Meta, Nvidia, Hugging Face, Mistral, and over 20 other firms, signed an open letter titled 'Open Weights and American AI Leadership.' The letter argues that U.S. leadership in AI should be judged by the strength of an open ecosystem rather than a single model. This approach is meant to foster innovation and reduce reliance on a few dominant providers. The letter also supports distillation, where smaller models learn from more capable ones, framing it as a continuation of open-source software traditions. It highlights that Chinese providers have faced criticism for using distillation, with the Trump administration reportedly planning actions against such models. This context likely explains the current push for open AI.

Microsoft's push for open AI aligns with its business model, which relies on cloud infrastructure and ecosystem lock-in through Windows and Office. By promoting more models on Azure, the company aims to keep customers within its platform, reducing reliance on costly models from OpenAI and Anthropic. This strategy could improve margins and prevent any single AI lab from becoming a major threat to Microsoft's cloud and operating system businesses. CEO Satya Nadella emphasized the importance of using 'the right model for each task' to cut costs. He previously warned that a few dominant AI models could capture the value of entire industries.

Microsoft is also replacing OpenAI and Anthropic models in products like GitHub Copilot, Excel, and Outlook with its own MAI family. However, independent benchmarks show MAI performing roughly at the level of Deepseek V3.2, raising questions about customer benefits. The company claims MAI performs as well or better than some models but only compares it against smaller, less powerful versions like GPT-5.4 Mini, Haiku, and 'GPT-5.6,' without specifying which variants or reasoning modes are used. This approach is not standard for benchmarking, and Microsoft's 'superintelligence team' is aware of this.

Source: thedecoder