The exact thesis of this note is: the announcement is a compute-efficiency claim tied to a completed pretraining run, not evidence of model availability or an independently established performance lead.
Observed fact: Microsoft says MAI-Base-2 outperformed every open-weight model it tested on held-out evaluations. Observed fact: it reports 4–12x greater pretraining-FLOPs efficiency than the strongest open base model it tested, DeepSeek-V4-Pro-Base, and says it put the savings into training a larger model. Improved training compute efficiency can be reinvested in model scale rather than treated only as a cost reduction. Microsoft has publicly marked a completed run and attached a quantified comparison, making the efficiency claim a concrete research milestone.
The claimed efficiency could let Microsoft pursue larger training runs within a given compute budget, if the comparison holds beyond its own evaluations. The source is Microsoft’s account of its own tests; it does not supply independent validation here, establish access or release timing, or substantiate MAI-Base-3 training preparations. Whether Microsoft publishes evaluation details or an independently reproducible comparison. Further signals to watch include release/access information and evidence clarifying the comparison methodology. Since no prior New Runtime record is supplied, this starts a distinct analytical record for this specific completed training run and its efficiency claim, without asserting a link to an earlier record.