Microsoft has finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters. Why it matters: Microsoft is no longer just reselling a partner’s frontier models — it now has an in-house base model at true frontier scale, and it says it trained that model far more efficiently than its rivals.
October 15, 2026 — Microsoft AI disclosed the completed pre-training run in a Lab Notes post dated October 15.
What did Microsoft announce?
MAI-Base-2 is the successor to MAI-Base-1, the base behind the MAI-Thinking-1 reasoning model Microsoft announced in June. The new model is larger and sparser, uses mixture-of-experts layers throughout, and was trained on more data, including vision data. Microsoft says the model is now in post-training, with further results to come.
The run itself was substantial: 16,000 NVIDIA GB200 GPUs working for 38 days across August and September 2026. Microsoft reports average model FLOPs utilization of 21% and goodput — the share of time training actually progressed — of 89%.
How does MAI-Base-2 compare with other models?
On Microsoft’s own held-out evaluations, measured in bits-per-byte so different tokenizers do not skew the comparison, the company says MAI-Base-2 scored better than every open-weight model it tested. The comparison set included DeepSeek-V4-Pro-Base, which Microsoft calls the strongest open base model, plus Kimi K3, Qwen3.8 and MiMo checkpoints.
The headline claim is efficiency. Using scaling-law fits against DeepSeek-V4-Pro-Base, Microsoft estimates its training recipe is about 4.2 to 12.3 times more FLOPs-efficient, and roughly 6.3 to 12.7 times more efficient than the recipe used for MAI-Base-1. Those are Microsoft’s own estimates from its own evaluations, not independent results.
Separately, Business Insider reported on October 15 that Microsoft AI chief Mustafa Suleyman told employees at a Thursday town hall that the new base model performed ahead of the model behind OpenAI’s GPT-5.6 Sol in Microsoft’s private testing, while using about a quarter of the training compute. That comparison has not been independently verified, and OpenAI has not publicly confirmed it.
How did Microsoft train it more efficiently?
Microsoft credits three areas: scaling science, model-and-infrastructure co-design, and bigger data pipelines. A technique it calls source scaling — controlling how often smaller training runs repeat the same data — helped its small-scale experiments predict full-scale behaviour. Engineering work included a dropless mixture-of-experts design with fixed memory use and a scaled-up version of the Muon optimizer.
Microsoft also stresses provenance: pre-training used organic data rather than distillation from other models, which the company says will let it license the weights to trusted partners through Microsoft Foundry.
Why it matters
For enterprise buyers, the practical point is choice. If post-training delivers, Microsoft will be able to offer a frontier-class model it controls end to end — in its own products first, then to Foundry customers — instead of depending entirely on OpenAI. For the wider market, the efficiency claim is the part to watch: if a 2.8-trillion-parameter model can really be trained for a fraction of the usual compute, the cost curve for frontier AI shifts. Until independent evaluations land, treat the performance lead as a company claim — but the completed training run, at this scale, is a fact.
Frequently asked questions
Is MAI-Base-2 available to use now?
No. Microsoft says the model is currently undergoing post-training. It plans to use it in its own products first and license weights to trusted partners through Foundry.
How big is MAI-Base-2?
It has 2.8 trillion total parameters, with 80 billion active per token — a sparse mixture-of-experts design trained on 16,000 GB200 GPUs over 38 days.
Did MAI-Base-2 really beat OpenAI’s model?
That is what Microsoft AI’s chief reportedly told employees, based on Microsoft’s private testing. It has not been independently verified.
Sources: Microsoft AI, Business Insider