Microsoft (Microsoft) announced yesterday local time that it has completed the pre-training of its new generation of SOTA base model MAI-Base-2. This model outperforms all the open-source weight models tested.
MAI-Base-2 The total parameter scale reaches 2.8T, with an active parameter scale of 80B. It was trained over 38 days on 6,000 NVIDIA GB200 GPU. Throughout the training process, the model's average floating-point computing power utilization rate was 21%, and the effective throughput was 89%.
According to Microsoft, the pre-training computing efficiency of MAI-Base-2 is 4 to 12 times that of DeepSeek-V4-Pro-Base.
Microsoft stated that three measures are particularly important for the successful pre-training of MAI-Base-2: scientific scaling, model-infrastructure collaborative design, and data scaling. While refining MAI-Base-2, this tech giant has already begun preparations for MAI-Base-3.
Related Reading:
Defeat OpenAI in the field where OpenAI excels: Microsoft claims its self-developed model surpasses similar products with only 1/4 of the computing power used by the latter
refer to
https :// microsoft.ai / news / frontier-pre-training-at-microsoft-ai /
