Pretraining is finished. Post-training is not
Microsoft AI's lab note says pretraining of MAI-Base-2 is complete. It is a mixture-of-experts base model with 80 billion active parameters and 2.8 trillion total. They call it state-of-the-art. The page date is October 15, 2026. The machine time is 16:09 UTC, 12:09 a.m. in Beijing on October 16. Post-training is still underway. Preparations for MAI-Base-3 have started. That is a plan, not a second release.
They say owning the stack lets them tune models and infrastructure together, first in their own products and then for Foundry customers. Post-training is not finished, so this is not a delivery note.
[1]
The 4-12x figure is their estimate against one base model
They write that on their held-out evaluations, MAI-Base-2 outperforms every open-weight model they tested. Against DeepSeek-V4-Pro-Base, which they call the strongest open base model, it is 4-12x more pre-training FLOPs-efficient. They put that saving into a larger model. They also write that DeepSeek-V4-Pro-Base is the only released base pretraining checkpoint with more than one trillion parameters, and that it was still trained on significantly fewer FLOPs than MAI-Base-2.
The held-out metric is bits per byte, lower is better. The prose says MAI-Base-2 is lower than the other models they tested. The chart's exact marks are not in the prose, so they are not filled in here. The comparison set is checkpoints available before October 2026. For Kimi K3, Qwen3.8-2.4T-A95B, and MiMo-V2.6-Pro-RL, only post-trained checkpoints are available. They note that post-training may affect these evaluations.
[1]16k GB200 GPUs, 38 days
The run used 16k GB200 GPUs over 38 days in August and September 2026. Average MFU, model FLOPs utilization, was 21 percent. Goodput, the fraction of time training progressed, was 89 percent. MFU is normalized by 2.5 PFLOPS, which they call the GB200 theoretical maximum in BF16, and it excludes FLOPs for rematerialization and the optimizer.
They write that after addressing memory fragmentation from dynamic routing and the overhead of finer-grained experts, MoE-layer MFU rose from 24 percent to 35 percent, and they used MoE in every layer. Muon was more stable than AdamW and, in their words, brought large efficiency gains. This post does not give the factor. On data, they write that the path is their own data, not distillation from other models. Unique training tokens grew by roughly 36 percent from MAI-Base-1 to MAI-Base-2. The web knowledge cutoff is March 2026.
[1]