Overview
- The Microsoft AI Superintelligence Team announced the seven MAI models at its Build developer conference on Tuesday, June 2, and placed MAI-Thinking-1 into private preview on Microsoft Foundry.
- MAI-Thinking-1 is presented as the flagship reasoning model, while MAI-Code-1-Flash, a 5-billion-parameter coding model, is rolling out into Visual Studio Code and GitHub Copilot.
- Microsoft emphasized that the models were trained from the ground up with ‘zero distillation,’ a process that would otherwise use another model’s outputs to train a new model, and said the training uses commercially licensed data to offer clear provenance for enterprise use.
- The company cited blind human tests and benchmark results to compare MAI-Thinking-1 with Anthropic’s Claude variants and reported high scores from third-party trackers such as Scale Labs, but external leaderboards like Arena.AI show mixed rankings and comprehensive independent verification is still pending.
- The move marks a strategic shift from Microsoft’s long reliance on partners such as OpenAI and Anthropic toward a hybrid approach of hosting partner models while building proprietary systems, a change that could reshape enterprise buying decisions and partner dynamics.