Microsoft has detailed MAI-Base-2, a base model with 2 trillion parameters that activates 80 billion per prompt. The company says it outperformed DeepSeek-V4-Pro-Base, described as the most capable open-source base model, on internal tests covering science, programming and math. By certain measures, the new model is up to 12 times more hardware-efficient. For business, this signals cheaper large-model inference if the efficiency holds in production systems.
Base model design and Azure licensing plan
MAI-Base-2 is not a finished large language model but a prototype that requires additional training rounds. Developers usually add reasoning capabilities and safety guardrails at that later stage before deployment. Microsoft plans to license the weights to trusted partners through the Azure cloud platform rather than release them openly. It will also use MAI-Base-2 as the foundation for a new series of language models, with more information promised in the near future.
The model uses a mixture of experts architecture, meaning only a subset of artificial neurons processes each prompt. Out of 2 trillion parameters, 80 billion activate per request, which reduces compute demand compared with activating the full network. Artificial neurons are organized into clusters called layers, and Microsoft changed the layer mix in this generation. While the previous base model combined mixture-of-experts layers with less efficient FFN modules, MAI-Base-2 consists entirely of mixture-of-experts layers.
Training ran for 38 days on 16,000 Nvidia GB200 chips, using a dataset of 50 trillion tokens. About a fifth of those tokens trained an integrated computer vision module that improves handling of tasks such as understanding scientific charts. That multimodal element distinguishes the project from a text-only base model effort. To stabilize training, Microsoft applied gradient clipping, which prevents errors that occur when gradients change while passing through model layers, and used Muon, a popular open-source AI tool, to speed the workflow.
What efficient base models mean for AI buyers
For companies building on AI, a more hardware-efficient foundation can lower inference costs and expand deployment options on Azure. Tasks involving code, math and scientific documents look like the first beneficiaries, given the benchmark focus and the vision module for charts. Small firms could gain access to stronger capabilities without operating their own training infrastructure, working through licensed models on cloud. Large organizations could standardize on derivatives of one base model across departments while keeping data and controls inside Azure.
The practical limits are clear: this is a base model, so no team can deploy it directly without post-training for reasoning, safety and domain fit. Access is restricted to trusted partners, with no public weights, pricing or timeline for wider availability disclosed. Performance claims rest on Microsoft internal benchmarks against DeepSeek-V4-Pro-Base, not independent evaluation. Buyers should ask what efficiency metric improved 12 times, how derivatives will be licensed, and what guardrails and evaluation results accompany each release.
The marker to watch is Microsoft disclosure about the new series of language models built on MAI-Base-2. Specifications, Azure availability dates and independent test results will show whether laboratory efficiency converts into lower cost per task. If partners ship products with documented gains on programming and chart-heavy documents, the model family will matter for procurement. Without that follow-up, it remains a research milestone.
