Microsoft has been working with OpenAI’s models for years. They built Azure OpenAI, invested billions in the partnership, and for most of the enterprise world, “Microsoft AI” and “GPT-4” became practically synonymous. That era is not ending, but it is evolving. Last week, Microsoft publicly detailed MAI-Base-2, its own internally trained frontier language model, and the technical ambition behind it is hard to overstate.
MAI-Base-2 is a 2.8-trillion-parameter model. Parameters are the numerical values a language model adjusts during training to learn how language works. A larger count generally correlates with more capable reasoning, though raw size is not everything. Microsoft uses an architecture called mixture-of-experts, or MoE, where only a fraction of those parameters activates for each request. In this case, 80 billion active parameters do the work per query. The rest sit ready but dormant. That design keeps inference costs manageable while still preserving the knowledge packed into a much larger model.
The company says MAI-Base-2 outperformed every open-weight model in its held-out pre-training evaluations. Open-weight models, like Meta’s Llama series or Mistral, are models whose weights are publicly available for anyone to run or modify. They have become serious competition to paid APIs. The same week this announcement landed, data from Vercel’s AI Gateway showed open-weight models handling over 70% of all tokens processed across the platform. Microsoft sees exactly how the landscape is shifting.
What matters most for enterprise customers is where this model goes. Microsoft plans to offer MAI-Base-2 through Azure AI Foundry, its platform for enterprise AI deployment. Foundry is where companies build, deploy, and manage AI applications with Microsoft-level compliance and data residency commitments. Having a Microsoft-trained model there means enterprise customers can access frontier-tier AI without a direct dependency on OpenAI’s pricing decisions, availability, or policy changes.
The training effort was no small undertaking. Microsoft ran MAI-Base-2 across 16,000 GPUs and built custom infrastructure to manage token routing across the model’s many expert layers without degrading efficiency. They also adopted an optimizer called Muon, which delivered more stable training dynamics than the AdamW optimizer that most teams use as a default. These are not details of a research experiment. This is a sustained, full-stack AI training investment with clear commercial intent.
For businesses evaluating AI strategy, the practical question is whether to build on a single vendor’s model or plan for portability. Microsoft’s investment points toward Azure becoming a complete AI stack, model included. That is worth factoring into platform decisions you are making today.
Want to explore how enterprise AI platforms could benefit your business? Let’s talk.