Microsoft Finishes Pre-Training a 2.8 Trillion-Parameter AI Model — Its Most Efficient Yet

By Hamza Faisal

Microsoft has finished building the largest AI model in its history — and its pitch is not raw size, but how little compute it took to get there. The company announced on October 15 that it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters, of which only 80 billion activate for any given token.

It is a big step up from MAI-Base-1, introduced in June with roughly 1 trillion total parameters and 35 billion active. The new model comes from Microsoft AI, the unit run by chief executive Mustafa Suleyman, as Microsoft pushes to own more of its foundation-model pipeline instead of leaning entirely on OpenAI.

16,000 Nvidia chips, 38 days

Microsoft trained MAI-Base-2 on 16,000 Nvidia GB200 GPUs over 38 days across August and September, processing about 50 trillion tokens of text and visual data with a knowledge cutoff of March 2026. The company reports 21% average model FLOPs utilization and 89% goodput — the share of training time that actually made progress — and says its pool of unique training tokens grew roughly 36% since MAI-Base-1.

The model is also Microsoft's first trained to see from the start. Instead of bolting on a separate image encoder, MAI-Base-2 learned from pictures and text together, from chart reading and document analysis to medical images. Vision-language data made up 18% of the training mixture.

The efficiency claim, and its caveats

The headline number: Microsoft estimates MAI-Base-2 is 4.2 to 12.3 times more compute-efficient to train than DeepSeek-V4-Pro-Base, which it calls the strongest open base model, and says it outperformed every open-weight model it tested on held-out evaluations.

The figures come with fine print. They are Microsoft's own estimates, fitted from scaling laws on selected evaluations rather than head-to-head results against every leading proprietary model — and the company acknowledges that for some comparison models, only post-trained checkpoints were available, which can shift results. The efficiency lead is real on Microsoft's chosen tests, but independent verification is still to come.

Microsoft also stresses what the model did not do: learn from other companies' systems. It says the training data is "organic" — every token processed in-house, with no distillation from rival models, and a proprietary web crawler that respects robots.txt.

Licensed, not released

MAI-Base-2 is now in post-training, the stage that turns a raw base model into something that follows instructions and reasons. Microsoft does not plan to release the weights publicly; instead, it will license them to trusted partners through Microsoft Foundry, its enterprise AI platform — a bet that companies will pay for a model whose training-data provenance is fully documented. Work on MAI-Base-3 has already begun.

The announcement sharpens Microsoft's push for AI independence. Business Insider has reported that Suleyman privately compared MAI-Base-2 against OpenAI's models, and the new efficiency figures double as a statement of intent: own the full stack, tune models and infrastructure together, and offer enterprises a cheaper path to customized intelligence. The next test is the practical one — whether the training gains translate into a model that writes code, reasons through hard problems and does real work.

By Hamza Faisal