Article 289: Microsoft Finishes Pretraining MAI-Base-2, a 2.8 Trillion-Parameter Model

Microsoft has completed pretraining MAI-Base-2: 2.8 trillion total parameters with 80 billion active in a mixture-of-experts layout, trained on 50 trillion tokens across 16,000 Nvidia GB200 GPUs in 38 days. Microsoft claims a 4x to 12x pretraining FLOPs-efficiency advantage over DeepSeek-V4-Pro-Base. The model ships to trusted partners through Foundry - it is not being released as open weights.

Key Takeaway: Note what Microsoft is claiming and what it is not. The number is about training cost, not finished capability, and there are no published benchmark results yet. The strategic signal is the delivery channel: a 2.8T model that goes to selected partners via Foundry is Microsoft selling the ability to make frontier models, rather than the model itself, which is precisely the position it occupies with its own customers already. For everyone else, the practical question is whether an efficiency claim of that size survives contact with fine-tuning and serving costs - history says the headline ratio rarely does.

Sources: Microsoft, AIToolsRecap, 17 October 2026