Microsoft AI published lab notes on MAI-Base-2, an MoE base model with 80 billion active and 2.8 trillion total parameters, pre-trained on 16,000 GB200 GPUs over 38 days in August and September 2026. The run averaged 21% MFU, normalised against GB200's 2.5 PFLOPS BF16 peak and excluding FLOPs for rematerialisation and the optimizer, with 89% goodput, meaning training progressed 89% of the time. Base models are compared on held-out bits-per-byte, with code evaluated on Microsoft's own codebase and STEM, math and Q&A sets commissioned from vendors; Microsoft says all evaluation sets were decontaminated against its training data. Against DeepSeek-V4-Pro-Base it estimates 4.2-12.3x better pre-training FLOPs efficiency, obtained by fitting independent scaling laws of BPB against training FLOPs on four evaluations and finding where a model from its own scaling ladder would match the reference; the gain over MAI-Base-1 is put at 6.3-12.7x. The recipe changes include "source scaling", which fixes the ratio of unique data to parameters on smaller ladder rungs by looping over a subset of the full 50-trillion-token run, joint learning-rate and weight-decay sweeps, improved MoE and Muon implementations, and activation gradient clipping. MAI-Base-2 is still in post-training.
Our takeThe efficiency claim rests on a scaling-law fit that includes the final model in the fit, so the matching point sits inside the fitted range and its width depends on how well the ladder transfers across a 2.8-trillion-parameter jump; quoting 4.2-12.3x is a fair way to show that spread. Read the MFU figure carefully: excluding rematerialisation and optimizer FLOPs puts real hardware utilisation below 21%. Disclosing 89% goodput at all is unusual and more useful than most frontier announcements, because it separates algorithmic efficiency from a cluster that simply stayed up.