Microsoft AI finishes MAI-Base-2 pre-training: 2.8T MoE, claims up to ~12× efficiency vs DeepSeek base
On October 15, 2026 Microsoft AI said it completed pre-training MAI-Base-2, an MoE base with 80B active and 2.8T total parameters now in post-training, claiming roughly 4–12× pre-training FLOPs efficiency versus DeepSeek-V4-Pro-Base and a path to license weights via Foundry.
Microsoft AI published the Lab Notes post “Frontier Pre-training at Microsoft AI” on October 15, 2026, saying it has finished pre-training MAI-Base-2. It describes an MoE foundation model with 80 billion active and 2.8 trillion total parameters; in June it had announced MAI-Thinking-1, a general-purpose model post-trained on MAI-Base-1.
Its accompanying PDF says MAI-Base-2 was trained on about 16,000 GB200 GPUs over 38 days in August–September 2026 on a roughly 50T-token mixture with a March 2026 knowledge cutoff. Versus MAI-Base-1 (about 35B active / 1T total with alternating dense and MoE layers), MAI-Base-2 uses MoE in every layer and adds native vision via an encoder-free VLM path; the VLM corpus is about 4.5T unique tokens and, with repetition, about 18% of the mixture.
Microsoft AI says MAI-Base-2 outperforms every open-weight model it tested on held-out evaluations, and estimates about 4–12× (also stated as roughly 4.2–12.3×) pre-training FLOPs efficiency versus DeepSeek-V4-Pro-Base as the strongest open base reference. Those figures are company evaluations and scaling-law projections, not independent replications.
The company says pre-training is built on “organic” data without distillation from other models, which it argues will support licensing weights to trusted partners through Foundry. MAI-Base-2 is in post-training with no public weights, API, or pricing disclosed; preparations for MAI-Base-3 are already under way, the post says.
Sources
- Frontier Pre-training at Microsoft AI — Microsoft AI
- Frontier Pre-training at Microsoft AI (PDF) — Microsoft AI