Microsoft Launches MAI Models and Maia 300, Cutting AI Inference Costs by 18%
Updated
Updated · InfoWorld · Nov 12
Microsoft Launches MAI Models and Maia 300, Cutting AI Inference Costs by 18%
3 articles · Updated · InfoWorld · Nov 12
Summary
Microsoft has begun rolling out its own MAI foundation models through Azure Foundry, targeting lower token costs rather than top-end benchmark leadership as AI inference grows more expensive.
MAI-Thinking-1 plugs into existing Azure OpenAI APIs, while MAI tuning tools let customers build custom versions and cap token use to manage application budgets.
MAI-Cyber-1.1-Flash already handles 90% of tasks previously sent to larger models in Microsoft’s MDASH security platform and cuts operating costs 18% versus version 1.0.
Maia 300 underpins that push with 36% more compute, 2.5x higher high-bandwidth memory performance and about 2x better performance per watt than prior Maia hardware.
The launch expands Microsoft beyond hosting OpenAI and other third-party models, betting that “good enough” models plus cheaper inference will win enterprises facing rising AI bills.