Microsoft AI reveals details on MAI-Base-2 pretraining: its 2.8T parameter MoE model achieves FLOPs efficiency 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.
Microsoft AI announces completion of MAI-Base-2 pretraining: an MoE base model with 80 billion active parameters and 2.8T total parameters. In held-out evaluations, its BPB score is lower than all tested open-weight models, while its pretraining FLOPs efficiency is 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original