Microsoft AI’s MAI-Thinking-1, a 35B-active-parameter MoE reasoning model released in public preview on Foundry in August 2026 after its June Build debut, established the baseline for in-house reasoning performance with strong AIME and SWE-Bench results. Trader sentiment now centers on the October 15 completion of pre-training for its larger successor foundation, MAI-Base-2, which will undergo post-training to produce the next dedicated thinking model. Microsoft’s independent “hill-climbing” pipeline, expanding Maia inference hardware, and competitive positioning against models like Claude and GPT variants add momentum, while timelines for any MAI-Thinking-2 release or Base-3 start remain fluid amid typical post-training cycles.