Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (IT Home · Oct 16, 2026)

Microsoft Completes 2.8T-Parameter MAI-Base-2 Pretraining

KEY FACTS

  • Microsoft announced it has completed pretraining of its next-generation foundation model, MAI-Base-2.
  • The model has 2.8T total parameters and 80B active parameters.
  • Training used 6000 NVIDIA GB200 GPUs over 38 days.
  • Microsoft says its performance beats all open-weight models tested.
  • Microsoft has already begun preparations for the next generation, MAI-Base-3.

KEY DATA

2.8TTotal Parameter Count
80BActive Parameter Count
6000 NVIDIA GB200Training GPU Count
38 daysTraining Duration

PHYSIX OBSERVATION

Microsoft has raised the foundation-model bar another notch with 2.8T parameters and a 6000-GPU cluster, and its emphasis on compute efficiency being several times that of rivals shows big-tech competition has shifted from piling on parameters to competing on training efficiency. For the industry, the gap between open-weight models and closed-source SOTA may keep widening, and the cost for smaller players to keep up is higher.

Source: IT Home report