MAI-Base-2 Claims 12x Efficiency, No Release

This is an unusually detailed lab note, but the headline efficiency figure is still a vendor-derived estimate—not a controlled head-to-head result or a usable model release.

• –Microsoft fitted scaling laws across internal runs to estimate the 4.2–12.3x advantage, rather than directly reproducing DeepSeek’s training recipe.
• –The published run details imply roughly 2.4–2.8 × 10²⁵ FLOPs, above the EU’s 10²⁵-FLOP systemic-risk threshold; that compute estimate is independent arithmetic, not a Microsoft-reported total.
• –MAI-Base-2 combines 2.8T total parameters, 80B active parameters, MoE layers throughout, Muon optimization, native vision training, and roughly 50T tokens.
• –Comparisons are constrained by Microsoft’s own held-out evaluations, and several rivals are post-trained checkpoints rather than comparable base models.
• –The important developer question—quality after instruction tuning, tool use, serving cost, and access—remains unanswered while post-training continues. [OrcaRouter analysis](https://www.orcarouter.ai/blog/mai-base-2-release-date)