Microsoft Completes Training of 2.8T-Parameter MAI-Base-2
Microsoft says its new mixture-of-experts foundation model has 80 billion active parameters and may be licensed through Azure AI Foundry.
Microsoft has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters, according to a Microsoft AI update published on October 15. The company says the model was trained with strong data provenance and is intended for licensing to trusted partners through Azure AI Foundry.
A platform move, not a public release
MAI-Base-2 is not yet a downloadable model or a broadly available API product. Microsoft’s announcement describes a completed pre-training milestone and a planned route to enterprise access, rather than an immediate open release. That distinction matters: the model’s strategic value currently lies in Microsoft’s ability to control distribution, serving, and partner access through Foundry.
The architecture also reflects the industry’s continuing shift toward very large sparse models. With only a fraction of its total parameters active for each token, MAI-Base-2 is designed to pursue frontier-scale capacity without paying the full computational cost of a dense 2.8-trillion-parameter model on every inference step. Microsoft has not yet published a complete evaluation suite, pricing, context limits, or a firm customer-availability date.
Why it matters
Microsoft has invested heavily in external model partnerships while building its own model family for products and infrastructure. MAI-Base-2 suggests that the company wants a stronger bargaining position: an internally controlled foundation model that can be offered through the same enterprise channel used for third-party systems.
The important question is therefore not the headline parameter count alone. It is whether Microsoft can turn a pre-trained internal model into a reliable, economically competitive service with enough differentiation to influence enterprise buying decisions. Until benchmarks and access terms arrive, claims about frontier performance remain unproven.