Microsoft Says Its New Base Model Trains Far Cheaper

Microsoft AI has finished pre-training a new foundation model called MAI-Base-2, and the headline it wants you to take away is not the size but the bill. In a post on its research site dated 15 October, the lab says the model is between four and twelve times more efficient per unit of pre-training compute than DeepSeek-V4-Pro-Base, which Microsoft treats as the best open base model to measure against.

The model itself is large. It uses a mixture-of-experts design with about 80 billion active parameters drawn from a 2.8-trillion-parameter total, so only a fraction of the network fires on any given token. That is the same trick most frontier labs now lean on to keep a very big model affordable to run.

What the efficiency claim means

Pre-training is the expensive part of building a model, the long run where it reads most of its data. If Microsoft really is getting four to twelve times more capability out of each FLOP spent there, it either saves a fortune or, as the company says it chose to do, spends the same money on a bigger model. Microsoft credits three things: better use of scaling laws to plan the run, hardware and model designed together, and more data. It also makes a point of saying the training data is organic rather than distilled from a rival model, a quiet jab at the common practice of training on another system's outputs.

The run took roughly 38 days, according to StreetInsider, on a cluster of Nvidia GB200 chips. Here the reports stop agreeing: some put the cluster at around 6,000 GPUs, others at 16,000. That gap matters for judging the efficiency claim, so it is worth flagging rather than papering over.

How much to trust the numbers

These figures are Microsoft's own, produced by its scaling-law comparison against a single reference model, and no outside group has checked them. A four-to-twelve-times range is also wide enough to invite scepticism about the method. The direction of travel is clear enough though. After building its own inference chip and a string of in-house models, Microsoft is trying to stop being the company that mostly ships other people's AI.

One thing MAI-Base-2 is not is open. The model follows MAI-Base-1 from June and will go to trusted partners rather than onto a download page, which puts it on the opposite side of the fence from the open-weight trillion-parameter models other labs have been releasing. For now you will have to take Microsoft's word for the savings.