Microsoft Unveils MAI-Base-2 Foundation Model with 2.8T Parameters

Key summary

Microsoft has introduced MAI-Base-2, a new foundation model designed for pre-training before being adapted into specialized models via post-training. The model utilizes a Mixture-of-Experts (MoE) architecture with a total parameter count of 2.8 trillion and an active parameter count of 80 billion, representing an increase from the previous MAI-Base-1 which had 1 trillion total parameters and 35 billion active parameters. Microsoft states that scaling up was necessary to match competitors in data volume, data quality, and compute resources. The model was trained using approximately 16,000 GB200 units over a period of 38 days. According to the company, MAI-Base-2 outperforms MAI-Base-1 on benchmarks and surpasses open-weight base models from China such as DeepSeek-V4-Pro-Base, as well as post-trained open-weight models including Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL. Microsoft attributes the high quality and lower cost to controlling the entire model pipeline and owning its infrastructure. Models derived from MAI-Base-2 will be integrated into other Microsoft products and made available for rent to customers.