IT之家 · 智能时代 · 10/16/2026, 11:34:25
Microsoft Completes Pre-training of 2.8T-Parameter MAI-Base-2 Model
Microsoft has completed pre-training its next-generation foundation model, MAI-Base-2, featuring 2.8T total parameters (80B active) and outperforming all tested open-weight models. The company claims a 4x to 12x improvement in compute efficiency over DeepSeek-V4-Pro-Base, attributed to scientific scaling, co-design, and data scaling strategies. Preparations for the successor, MAI-Base-3, are already underway.
SOURCE COVERAGEFull Original Report
On October 16, Microsoft announced yesterday (local time) that it has completed the pre-training of its next-generation state-of-the-art (SOTA) foundation model, MAI-Base-2. This model outperforms all open-weight models included in the benchmark suite.
MAI-Base-2 features a total parameter count of 2.8T with 80B active parameters, trained over 38 days on 6,000 NVIDIA GB200 GPUs. Throughout the training process, the model achieved an average floating-point utilization rate of 21% and an effective throughput of 89%.
According to Microsoft, the pre-training compute efficiency of MAI-Base-2 is 4–12 times higher than that of DeepSeek-V4-Pro-Base.
Microsoft stated that three initiatives were particularly critical to the successful pre-training of MAI-Base-2: scientific scaling, model-infrastructure co-design, and data scaling. While refining MAI-Base-2, the tech giant has already begun preparations for MAI-Base-3.