Microsoft Completes Pre-Training of 2.8T-Parameter MAI-Base-2
MAI-Base-2 was trained for 38 days on 16,000 Nvidia GB200 GPUs and Microsoft says it used data it collected and processed itself, not distillation.
Microsoft announced on October 15, 2026 that it completed pre-training of MAI-Base-2, its self-developed large AI model with 2.8 trillion parameters, of which 80 billion are activated per token. The model uses a Mixture-of-Experts (MoE) architecture.
Microsoft said MAI-Base-2 outperformed all open-weight models used as comparison targets in its internal evaluation. It also said the model achieved 4.2 to 12.3 times higher efficiency in the floating-point operations (FLOPs) required for pre-training than DeepSeek's DeepSeek-V4-Pro-Base. Compared with MAI-Base-1, released in June 2026, compute efficiency improved by 6.3 to 12.7 times, according to the company.
MAI-Base-2 was trained over 38 days in August-September 2026 on 16,000 Nvidia GB200 GPUs. Training data grew by about 36% in unique tokens after deduplication compared with MAI-Base-1. Microsoft applied MoE to all layers and raised the model compute utilization rate of MoE layers from 24% to 35%, and used the Muon optimizer and a gradient-clipping technique to improve training stability. The company said it processed all training tokens internally using its own collected and processed data rather than distillation from other AI models.
Microsoft said MAI-Base-2 was designed from the start to train on images and text together, with vision-language model (VLM) data accounting for 18% of the total 50 trillion-token training composition.
Internal comparison and next steps
Mustafa Suleyman, CEO of Microsoft AI, reportedly claimed at an employee town hall that Microsoft built a stronger pre-trained model than the base model of GPT-5.6 Sol using about one quarter of the compute OpenAI used, according to Business Insider. The comparison with OpenAI was based on a private evaluation, and the public technical report does not provide detailed results to verify it.
Microsoft plans to use MAI-Base-2 as the foundation for its next-generation in-house models, applying it first to its own products and later expanding supply to enterprise customers, with model weight distribution planned through its AI platform Foundry. MAI-Base-2 is currently in the post-training stage, and Microsoft said preparation for the follow-up model MAI-Base-3 has already begun.
Quick answers
How many parameters does Microsoft's MAI-Base-2 have?
MAI-Base-2 has 2.8 trillion parameters, of which 80 billion are activated per token, and it is based on a Mixture-of-Experts architecture.
What hardware was used to train MAI-Base-2?
Microsoft trained MAI-Base-2 for 38 days in August-September 2026 using 16,000 Nvidia GB200 GPUs.
When will MAI-Base-2 be available to customers?
Microsoft said MAI-Base-2 will first be applied to its own products and later expand to enterprise customers, with weight distribution planned through its AI platform Foundry. The model is currently in the post-training stage.