Microsoft's new MAI models
Microsoft launched two small MAI LLMs trained on clean, commercially licensed data without third-party distillation.
“We trained [MAI-Thinking-1] from the ground up on enterprise grade, clean and commercially licensed data, without distillation from third-party models.”
Microsoft announced two new in-house text models, MAI-Thinking-1 (35B reasoning) and MAI-Code-1-Flash (5B, for GitHub Copilot), notable for their unusually low parameter counts and a claim that MAI-Thinking-1 is preferred over Sonnet 4.6 in blind evaluations. The most significant signal is Microsoft's emphasis on training end-to-end on clean, appropriately licensed data without distilling from other models, which could position them as a differentiated player in the IP-clean model space.