Microsoft unveils MAI-Base-2 AI model and StudentSim educational research

Microsoft has announced significant advancements in artificial intelligence, including a new foundational model and a specialized research system for educational AI.

Microsoft AI CEO Mustafa Suleyman revealed the development of MAI-Base-2, a pre-trained model utilizing a Mixture-of-Experts (MoE) architecture with 2.8 trillion total parameters. Suleyman stated that the model outperforms OpenAI’s GPT-5.6 Sol in pattern recognition and information prediction tasks, while requiring only one-quarter of the computing power used by OpenAI’s equivalent models. This development signals Microsoft’s increasing efforts to reduce its reliance on OpenAI, despite its massive investment in the partner.

Separately, Microsoft Research introduced StudentSim, a system designed to evaluate AI tutoring tools. Rather than simply checking for correct answers, StudentSim uses AI-driven student simulators to test whether a tutor can effectively guide a student to correct their own mistakes. In chess-based testing, StudentSim demonstrated higher behavioral fidelity and guidance responsiveness compared to prompt-driven GPT-5.4 models. The research aims to shift AI educational standards from measuring output accuracy to measuring actual student learning outcomes.