Microsoft AI's first 1T+ pre-training run, disclosed as lab notes rather than a model launch. The October 15 post describes MAI-Base-2 as "an MoE foundation model with 80 billion active and 2.8 trillion total parameters" — nearly triple the 1T total of MAI-Base-1, the base architecture behind MAI-Thinking-1. Microsoft says it "trained MAI-Base-2 on 16k GB200 GPUs over 38 days in August and September 2026," with an average MFU of 21% (normalized by 2.5 PFLOPS, the GB200's theoretical BF16 maximum) and 89% goodput. The training mixture is 50T tokens; the corpus "grew by roughly 36% from MAI-Base-1 to MAI-Base-2, measured in unique training tokens," and Microsoft repeats the line that distinguished the MAI program from the start: "Our pre-training is built on organic data which does not involve distillation from other models."

What changed in the recipe. Where MAI-Base-1 alternated MoE with dense layers, Microsoft says it "adopt[s] MoE in every layer for MAI-Base-2," with a multi-round dropless MoE that "achieves fixed memory usage under arbitrary routing imbalance" and lifted MoE-layer MFU from 24% to 35%. Muon replaces AdamW — it "delivered more stable training dynamics and large efficiency gains over AdamW," at a cost of 2.1% (165 ms) of the average training iteration time for the distributed Newton–Schulz orthogonalization. Vision is native rather than bolted on: the model "is pre-trained natively with vision using an encoder-free VLM architecture," and "VLM data accounted for 18% of the 50T-token training mixture." For stability at this scale Microsoft fits scaling laws for learning rate, learning-rate×weight-decay and dropout ("These laws defined MAI-Base-2's hyperparameters") and clips activation gradients "at selected places in the network" rather than only parameter gradients.

The headline claim is an efficiency ratio, not a score. Microsoft says the recipe is "4.2–12.3x more FLOPs-efficient" than the strongest open base model it benchmarked, DeepSeek-V4-Pro-Base, and that "MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations." No absolute benchmark numbers appear anywhere in the post — the comparisons are loss and held-out curves on Microsoft's own internal evaluations, so nothing here is independently checkable.

Status: this is a base model still in post-training, and there are no weights. Microsoft states plainly that "MAI-Base-2 is currently undergoing post-training, and we look forward to sharing more results soon," and that the eventual plan is to "license our model weights to trusted partners through Foundry" — i.e. partner licensing, not a public release. Also undisclosed: layer and expert counts, context length, precision beyond the BF16 MFU normalization, and any release date. The post carries no named authors.

Model Details

Architecture MOE
Parameters 2.8T
Active params 80B
Training tokens 50T
Training hardware 16K GB200
Training time 38 days
Optimizer Muon
License Proprietary
frontiermoemultimodalpretrainingproprietary

Related