Microsoft Completes Pre-training of 2.8 Trillion Parameter MAI-Base-2
Microsoft has finished pre-training MAI-Base-2, a new AI model with 2.8 trillion parameters designed to outperform open-weight models with significantly lower computing requirements.
Microsoft completed the pre-training of MAI-Base-2, a new artificial intelligence model featuring 2.8 trillion total parameters. The model utilizes a mixture-of-experts architecture with 80 billion active parameters, which the company reports allows it to outperform all tested open-weight models while requiring 4.2 to 12.3 times less computing power than DeepSeek-V4-Pro-Base.
Training took place over 38 days during August and September using 16,000 Nvidia GB200 GPUs and a corpus of approximately 50 trillion tokens. The model includes native vision capabilities, with image-related data making up 18% of the training mixture. Microsoft trained the model on original data rather than distilled knowledge to facilitate future licensing through Azure AI Foundry.
While MAI-Base-2 is currently undergoing post-training, the company has not yet disclosed pricing or commercial availability dates. Microsoft has already begun preparations for a successor model, MAI-Base-3.
