Microsoft announced yesterday local time that its new-generation SOTA foundation model MAI-Base-2 has completed pre-training, and officially stated that its performance is stronger than all open-weight models tested.
These numbers are quite hardcore: MAI-Base-2 has a total parameter scale of 2.8T with 80B active parameters, and was trained in 38 days on 6000 NVIDIA GB200 GPUs. Throughout the training, the model achieved an average floating-point compute utilization of 21% and an effective throughput of 89%—holding steady at this efficiency on an extremely large cluster is itself a testament to engineering strength.
Compute Efficiency Leaves DeepSeek Behind by 4 to 12 Times
Even more intriguing is the efficiency comparison. Microsoft claims that MAI-Base-2's pre-training compute efficiency reaches 4 to 12 times that of DeepSeek-V4-Pro-Base. At a time when "burning money to build infrastructure" defines the industry's first half, opening up an order-of-magnitude gap in output per unit of compute carries more strategic significance than simply piling on parameters.
Microsoft attributes this success to three things: scientific scaling, co-design of model and infrastructure, and data scaling. In other words, rather than just adding more cards, the algorithm, hardware, and data recipe were polished in sync. Even while wrapping up MAI-Base-2, this giant has quietly begun preparations for MAI-Base-3—in the foundation model arms race, Microsoft clearly has no intention of falling behind, and is instead accelerating to the front of the pack.