微软完成 MAI-Base-2 预训练：2

Microsoft completed MAI-Base-2 pre-training: 2.8T parameters crushed open source opponents, and the computing power efficiency exceeded DeepSeek several times. Microsoft announced yesterday (local time) that the new generation SOTA basic model MAI-Base-2 has completed pre-training. Officials said that its performance is stronger than all open source weight models tested. This set of numbers is quite hard-core: MAI-Base-2 has a total parameter size of 2.8T, active parameters 80B, and it took 38 days to train on 6,000 NVIDIA GB200 GPUs. During the entire training period, the model's average floating-point computing power utilization was 21% and the effective throughput was 89% - being able to stabilize this efficiency on a very large cluster is a footnote to its engineering strength. The computing power efficiency is 4 to 12 times better than DeepSeek. What is more worth pondering is the efficiency comparison. Microsoft said that the pre-training computing power efficiency of MAI-Base-2 is 4 to 12 times that of DeepSeek-V4-Pro-Base. At a time when "burning money to build infrastructure" has become the first half of the industry, being able to widen the gap in output per unit of computing power by an order of magnitude is more strategic than simply stacking parameters. Microsoft attributes this success to three things: scientific scaling, co-design of models and infrastructure, and data scaling. In other words, it is not about adding cards blindly, but allowing algorithms, hardware and data recipes to be polished simultaneously. While finishing MAI-Base-2, the giant has quietly started preparations for MAI-Base-3 - an arms race for basic models. Microsoft obviously does not want to fall behind, but has accelerated to the front row. via AI News (author: AI Base)