Microsoft disclosed on October 15 that it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active per token. Post-training is now underway.
The announcement offers a detailed look at Microsoft’s effort to develop models and training infrastructure together, while leaving the finished product’s capabilities and availability for a later update.
Efficiency gains went into a larger model
Microsoft says the training run used 16,000 NVIDIA GB200 GPUs over 38 days in August and September. Its reported efficiency improvements came from work across data preparation, model architecture and infrastructure.
The company estimates that its training recipe would require 4.2–12.3 times less computation to match DeepSeek-V4-Pro-Base on selected pre-training evaluations. It says those savings were reinvested in training a larger model.
These are Microsoft’s estimates under its evaluation method. They should not be read as equivalent reductions in the price of running a finished assistant.
What the evaluations can tell us
The comparisons use held-out code, question-answering, mathematics and STEM material. Microsoft also notes that some comparison models were post-trained checkpoints, which can affect the results.
For readers, the distinction between prediction quality and task completion is important. A model may learn strong representations during pre-training, but a useful assistant must also interpret instructions, select tools appropriately and recover when an attempted action fails.
Those behaviors require their own evaluation. A lower pre-training loss cannot, by itself, establish that an agent will complete a software change or analyze a business document reliably.
The next test is the product
Microsoft plans to make model weights available to trusted partners through Foundry, but has not announced a general release date.
AIYXL Perspective: The disclosure provides evidence of progress in Microsoft’s underlying training pipeline. The next stage will show how that progress translates into usable models: which tasks improve, what access costs, and whether gains persist across customers’ actual workloads.