Microsoft Finishes Pre-Training MAI-Base-2 With 2.8 Trillion Parameters
The model Microsoft is describing scales past MAI-Base-1 and buys its gains partly from image data folded into pre-training, so the foundation layer for its next round of products is now multimodal rather than text-only.
Reporting from 1 source: GIGAZINE.
Microsoft says pre-training is complete for MAI-Base-2, a mixture-of-experts model with 80 billion active parameters and 2.8 trillion total. It was trained over 38 days from August to September 2026 on 16,000 GB200 GPUs, and unlike MAI-Base-1 it takes in image data, which makes up 18 percent of its 50 trillion training tokens. Post-training is underway and work on MAI-Base-3 has begun.
The training run itself is the number Microsoft chose to publish: 38 days across 16,000 GB200 GPUs. Microsoft says pre-training compute was cut to one twelfth while matching the pre-training score of DeepSeek-V4-Pro-Base. MAI-Base-2 also adds images at the pre-training stage, with image-related data at 18 percent of 50 trillion tokens, a change from MAI-Base-1. Microsoft says the model was not built by distillation from other models, and that the training pipeline now covers knowledge up to March 2026.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.