Microsoft Completes Pre-training of New AI Model MAI-Base-2
Microsoft announced it has completed the pre-training of its new foundational model, MAI-Base-2. The company reports the model outperforms all tested open-source models.
Microsoft announced recently that it has completed the pre-training of its next-generation foundational model, MAI-Base-2. According to the company, the model performs better than all open-source models it was compared against.
The MAI-Base-2 model has a total parameter scale of 2.8 trillion, with 80 billion active parameters. Its training was conducted on 6,000 NVIDIA GB200 GPUs over 38 days. During the training process, the model achieved an average compute utilization of 21 percent.
Microsoft states that the training compute efficiency of MAI-Base-2 is 4 to 12 times greater than that of DeepSeek-V4-Pro-Base. The company highlighted the importance of scientific scaling, co-design of the model and infrastructure, and data scaling for the successful pre-training.
Microsoft has already begun preparations for its next model version, MAI-Base-3. This development continues the company's efforts to enhance AI model performance and pursue advancements in the field.