Called Phi-4, the model improves in several areas over its predecessors, Microsoft claims, particularly in solving math problems. That’s partly the result of better training data quality.