TL;DR
Microsoft AI announced October 15 that it completed pre-training of MAI-Base-2: a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active per token, trained on 16,000 Nvidia GB200 GPUs over 38 days. Microsoft says its recipe is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base, the strongest open base model — but those are the company's own estimates, and the model isn't released: it's in post-training, with weights eventually licensed to trusted partners through Foundry.
What Microsoft announced
The details come from a Microsoft AI Lab Notes post, 'Frontier Pre-training at Microsoft AI,' published October 15, 2026. MAI-Base-2 is an MoE model: 2.8 trillion parameters total, with 80 billion active for any given token — larger and sparser than MAI-Base-1 from June. Microsoft trained it on 16,000 GB200 GPUs for 38 days across August and September 2026, reporting 21% average model FLOPs utilization and 89% goodput (the share of time training actually progressed).
Reporting on the post puts the training mix at about 50 trillion tokens with no distillation from other companies' models — Microsoft calls its data 'organic' — and a March 2026 knowledge cutoff. Unlike MAI-Base-1, MAI-Base-2 trained on vision from the start: about 18% of the mix is image and text data together, covering charts, captions, spatial reasoning, and medical images.
The efficiency claim, with caveats
Microsoft's headline number is compute efficiency: it says MAI-Base-2's recipe is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, the strongest open base model, and 6.3 to 12.7 times more efficient than MAI-Base-1. That comes from fitting scaling laws to four held-out evaluations, not from head-to-head released benchmarks — Microsoft's own method, on Microsoft's own evaluations. Take it as a company claim until independent testing exists.
The techniques behind it are the interesting part for builders: 'source scaling' to keep small-scale experiments honest about data repetition, MoE in every layer with a multi-round 'dropless' routing scheme that raised MoE-layer utilization from 24% to 35%, and the Muon optimizer instead of AdamW for more stable training at this scale. Microsoft says work on MAI-Base-3 has already started.
Not open, not available — and that's the point
MAI-Base-2 is a base model in post-training: no API, no weights, no release date. Microsoft's plan is to license the weights to trusted partners through Foundry — the enterprise pitch is owning the full stack: strong data provenance, no distillation from rivals' models, and co-designed infrastructure tuned for the lowest cost per useful token.
That puts Microsoft in a different lane from Meta's open-weight push the same week. Meta is giving weights away (with limits); Microsoft is building a proprietary stack to rent. Both claim the efficiency edge matters most. One of them will be right about where the money goes.
What to do with this
If you build on models: nothing changes this week — you can't touch MAI-Base-2. But if you train or fine-tune at scale, watch what Microsoft publishes next. Dropless MoE routing and Muon are already in the open literature; if Microsoft's 4x efficiency edge is real, expect those ideas to show up in open training recipes within months.
If you buy AI for a company: the Foundry licensing plan means Microsoft-trained weights may soon be another line item alongside open weights and API subscriptions — one with provenance Microsoft will stand behind. Worth knowing about before your next contract talks.