Microsoft Corp.'s AI research division announced MAI-Base-2, a new base large language model featuring 2 trillion parameters that delivers high efficiency and advanced capabilities, aiming to improve next-generation AI development.
- MAI-Base-2 uses MoE architecture activating 80 billion parameters of 2 trillion total
- Outperforms leading open-source base model on science, programming, and math tests
- Trained for 38 days on 16,000 Nvidia GB200 chips with 50 trillion token dataset
What happened
Microsoft revealed its latest artificial intelligence base model called MAI-Base-2, which contains 2 trillion parameters. It is not a fully trained large language model but rather a prototype base model that developers can further train to add reasoning and safety capabilities. The model employs a mixture of experts (MoE) architecture that only activates a fraction of its parameters for any given prompt, improving hardware efficiency significantly.
MAI-Base-2 outperformed the leading open-source base model, DeepSeek-V4-Pro-Base, on internal benchmarks related to science, programming, and mathematics. The model was trained over 38 days on 16,000 Nvidia GB200 chips using a dataset of 50 trillion tokens, about a fifth of which supported an integrated computer vision module to improve the model’s ability to understand scientific charts and visual data.
Why it matters
Microsoft’s MAI-Base-2 demonstrates a significant leap in large language model efficiency by using a fully MoE-based architecture, allowing it to activate only 80 billion parameters out of 2 trillion in response to tasks. This efficiency can reduce the compute power needed, lowering costs and enabling faster responses in practical AI applications. Such architecture advances are crucial as AI models grow larger and more complex.
By licensing MAI-Base-2’s weights to trusted partners through its Azure cloud platform, Microsoft is positioning itself to accelerate innovation for enterprise AI solutions. The vision module integration further broadens the model’s capability set beyond text to include complex visual reasoning, a key differentiator for scientific and technical AI applications.
What to watch next
Microsoft plans to build a new series of large language models powered by MAI-Base-2 and share further details about these developments soon. Industry observers will look for how Microsoft’s approach to MoE architectures influences future AI model designs and whether its efficiency claims translate into real-world cost savings and performance gains.
Additionally, the rollout of MAI-Base-2’s licensing via Azure may set a precedent in AI model distribution and collaboration with partners. Tracking Microsoft’s deployment strategy and the adoption of integrated vision AI within language models will be key to understanding the evolving landscape of enterprise AI capabilities.