Microsoft unveils MAI-Base-2, a 2.8T-parameter MoE base model
Microsoft said on October 15 that it has finished pretraining MAI-Base-2, a mixture-of-experts base model with 2.8T total and 80B active parameters, trained for 38 days on 6,000 Nvidia GB200 GPUs at 21% average floating-point utilization and 89% effective throughput. Microsoft claims its pretraining compute efficiency is 4x–12x that of DeepSeek-V4-Pro-Base and that the model outperforms all open-weight models it tested, and says work on MAI-Base-3 has begun. Microsoft AI CEO Mustafa Suleyman told an internal employee meeting that the model beats the OpenAI model powering GPT-5.6 Sol, codenamed Spud, while using about a quarter of the compute, a comparison he said was based on OpenAI technical information Microsoft holds; OpenAI has not commented. In a separate report the same day, Microsoft said a new SharePoint-based Work IQ knowledge engine lets Copilot agents respond about 4x faster with roughly 30% fewer tokens without changing the underlying model, available to usage-billed Frontier Preview tenants.
If Microsoft's performance and efficiency claims hold up, they would reduce its reliance on OpenAI models and push frontier competition further toward training and retrieval efficiency.