Microsoft Completes Training of MAI-Base-2, a 2.8 Trillion Parameter Foundation Model

Quick Facts

• MAI-Base-2 has 2.8 trillion total parameters and 80 billion active parameters, using a Mixture of Experts architecture trained entirely in-house on 16,000 NVIDIA GB200 GPUs over 38 days.
• Microsoft claims the model is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base and outperforms every tested open-weight model on held-out evaluations.
• Microsoft AI CEO Mustafa Suleyman told employees MAI-Base-2 is a stronger base model than the one powering OpenAI's GPT-5.6, built at roughly one-quarter the compute cost.

Microsoft completed pre-training of MAI-Base-2 on October 15, 2026, marking its largest foundation model to date. The announcement positions the model as a direct competitor to the strongest open-weight and proprietary models available.

MAI-Base-2 is a Mixture of Experts model with 2.8 trillion total parameters. It activates 80 billion parameters per prompt, routing each token through 7 of 512 expert networks plus one shared expert. Microsoft's technical blog says the entire network is built from MoE layers, replacing the hybrid design used in its predecessor MAI-Base-1, which mixed MoE layers with less efficient feed-forward network modules.

The model trained on a 50 trillion-token corpus with a March 2026 knowledge cutoff. That training set grew roughly 36% in unique tokens compared to MAI-Base-1. The run averaged 21% model FLOPs utilization and 89% goodput across the 38-day training window.

Vision capability is built in from the start. MAI-Base-2 uses an encoder-free architecture for vision-language tasks, with a 4.5 trillion token visual corpus accounting for 18% of the full training mixture. Tasks covered include image captioning, chart understanding, spatial reasoning, and medical image understanding.

Microsoft trained the model using a technique called source scaling, which makes small-scale experiments reliably predict large-scale outcomes. The method keeps the ratio of unique data to parameters fixed across a scaling ladder, preventing smaller runs from favoring choices that overfit at larger scale.

Benchmark results show MAI-Base-2 is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, the top open-weight base model, and 6.3 to 12.7 times more efficient than MAI-Base-1. Microsoft attributes the gains to data quality and modeling choices rather than raw compute scale.

Suleyman made more pointed claims in an internal town hall. He told employees MAI-Base-2 performs better than an internal OpenAI model he called Spud, which he said powers GPT-5.6 Sol. He said Microsoft trained its model at roughly one-quarter the compute cost OpenAI used for its comparable model. Those comparisons are based on private tests and technical information Microsoft holds about OpenAI systems. Microsoft has not published the results.

The remarks carry weight given the history between the two companies. Microsoft remains a major investor in OpenAI, holds a stake likely worth more than $100 billion, and sells OpenAI models through Azure. The relationship has grown more complicated as Microsoft expands its own model portfolio.

The business case for MAI-Base-2 runs through data provenance. Microsoft says its training pipeline uses organic data with no distillation from other models, processed entirely in-house. Its proprietary crawler follows robots.txt and excludes sources on the USTR Notorious Markets list. That chain of custody is designed to support enterprise licensing through Foundry, Microsoft's model platform for external customers.

Microsoft said owning the full stack from data through infrastructure lets it tune models and systems together, aiming for top performance at lower cost in its own products before extending access to Foundry customers. No public release date for MAI-Base-2 has been announced.