Report
Microsoft Completes Pre-Training of MAI-Base-2 Frontier Model
Microsoft AI says its model outperformed tested open-weight models on held-out evaluations and was 4–12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base.
TLDR
Microsoft AI said it had built pre-trained base models in the 3T-parameter weight class entirely in-house, with post-training results to follow. TestingCatalog quoted the company’s technical blog as saying MAI-Base-2 outperformed every open-weight model it tested on held-out evaluations and was 4–12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. TestingCatalog also said preparation for MAI-Base-3’s training cycle was underway.
Combined views
—
5 Sources, first seen ago
