Microsoft unveils MAI-Base-2, a 2-trillion parameter AI model built for hardware efficiency
The new base model uses a mixture-of-experts architecture to outperform rivals on internal benchmarks while using fewer resources.
Key points
- Microsoft’s MAI-Base-2 features 2 trillion parameters but activates only 80 billion per prompt.
- The model outperformed DeepSeek-V4-Pro-Base on internal science, math, and programming tests.
- Training required 16,000 Nvidia GB200 chips over 38 days using a 50 trillion token dataset.
Microsoft Corp. detailed its new artificial intelligence base model, MAI-Base-2, on Thursday, according to a report by SiliconANGLE. The company’s research group presented the algorithm as a significantly more efficient alternative to leading open-source options, highlighting its advanced architecture and performance capabilities.
MAI-Base-2 is not a finished large language model but a foundational prototype that developers must further train to add reasoning and safety features. Microsoft claims the model surpasses DeepSeek-V4-Pro-Base, currently the most capable open-source base model, on a series of internal benchmarks covering science, programming, and mathematics.
Technical Architecture and Training
The model utilizes a mixture-of-experts architecture, which activates only a subset of its artificial neurons to process prompts rather than the entire network. MAI-Base-2 contains 2 trillion parameters but engages just 80 billion of them during operation. This design, combined with using entirely mixture-of-experts layers instead of mixing them with less efficient modules, contributes to its high performance.
Microsoft stated that MAI-Base-2 is up to 12 times more hardware-efficient than competitors by certain measures. The training process spanned 38 days and employed 16,000 Nvidia Corp. GB200 chips. The dataset comprised 50 trillion tokens, with approximately one-fifth dedicated to training an integrated computer vision module that improves chart interpretation.
Development Methodology
Developers refine large language models by observing task completion and providing algorithmic feedback in the form of gradients. These mathematical objects travel through the model’s layers to recalibrate them and boost output quality. Microsoft noted that gradients change with each layer pass, which can sometimes cause bugs that disrupt training workflows under certain circumstances.
To prevent these issues from affecting MAI-Base-2, Microsoft employed a technique called gradient clipping. This method stabilizes the training process by managing the changes in gradients as they move through the neural network layers. The resulting base model serves as a starting point for creating specialized large language models with additional safety guardrails and reasoning capabilities.
What should you do?
Security and technology professionals should monitor vendor announcements regarding new AI base models to assess potential integration opportunities. Teams should evaluate how new mixture-of-experts architectures might impact their existing infrastructure requirements and hardware compatibility. Organizations should review their current data processing pipelines to ensure they can handle the scale of training datasets described in these developments.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Microsoft
- Monitor official vendor channels for updates on model availability and licensing terms.
- Assess current hardware infrastructure to determine compatibility with new AI architectures.
- Review data governance policies to ensure compliance when using large training datasets.
- Evaluate integration requirements for base models before committing resources to further training.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does MAI-Base-2 compare to other open-source models?
Microsoft says MAI-Base-2 outperformed DeepSeek-V4-Pro-Base on internal benchmarks for science, math, and programming tasks.
What is the parameter count of the new model?
MAI-Base-2 has 2 trillion parameters, though it only activates 80 billion of them when processing a prompt.
How was the model trained?
Microsoft trained the model for 38 days using 16,000 Nvidia GB200 chips and a dataset of 50 trillion tokens.
