Microsoft's new base model uses a mixture-of-experts design to process prompts with significantly less computing power than rivals.
Key points
- MAI-Base-2 activates only 80 billion of its 2 trillion parameters per prompt
- The model outperformed DeepSeek-V4-Pro-Base on internal science and math tests
- Training required 16,000 Nvidia GB200 chips over a 38-day period
Microsoft has released details on MAI-Base-2, a large language model base that prioritizes hardware efficiency. The company states the prototype significantly outperforms leading open-source alternatives in internal benchmarks.
What happened
Microsoft’s artificial intelligence research group detailed the algorithm on Thursday according to SiliconANGLE. The company described MAI-Base-2 as a base model rather than a finished large language model. Developers must apply additional training rounds to add reasoning features and safety guardrails before deployment. Microsoft claims the new model is up to 12 times more hardware-efficient than competitors by certain measures. It surpassed DeepSeek-V4-Pro-Base, the current leading open-source base model, on internal tests involving science, programming and math questions.
The architecture relies entirely on mixture-of-experts layers instead of mixing them with less efficient feed-forward neural modules. This design allows the system to activate only a subset of its artificial neurons for each prompt. The model contains 2 trillion parameters but activates just 80 billion during operation. Microsoft trained the system for 38 days using 16,000 Nvidia GB200 chips. The training dataset included 50 trillion tokens, with one-fifth dedicated to a computer vision module. This addition improves the model’s ability to interpret scientific charts compared to its predecessor. The team used gradient clipping to prevent bugs caused by changing mathematical gradients during training.
Why it matters
Base models serve as foundational prototypes for developers. They require further refinement to become usable applications. Efficiency in these models reduces the computing resources needed for training and inference. This shift allows organizations to run advanced AI tasks with lower hardware costs. Understanding the difference between base models and finished products helps readers evaluate vendor claims. It also highlights the importance of architectural choices like mixture-of-experts in scaling performance. Teams usually assess these models based on benchmark results and resource requirements before integration.
The fix
No security vulnerability or software bug was reported in this announcement. Therefore, no technical fix or patch is required for this event.
Why it matters
Organizations evaluating new AI models should verify independent benchmark results. They must also assess the computational costs of training and running these systems. Understanding architectural differences helps in selecting models that fit specific hardware constraints. This approach ensures that efficiency claims translate into actual operational benefits.
What to do and how to stay safe: Microsoft
- Review internal hardware capacity before testing new large language models
- Verify performance claims against independent third-party benchmarks
- Assess the specific training requirements of base models
- Monitor vendor announcements for additional safety training details
Step-by-step guide: What is password policies and how they protect systems
General security guidance from the Crypt Weekly newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is MAI-Base-2?
It is a base model prototype developed by Microsoft that uses a mixture-of-experts architecture.
How does it compare to other models?
Microsoft states it outperformed DeepSeek-V4-Pro-Base on internal science and math benchmarks.
What hardware was used for training?
The model was trained using 16,000 Nvidia GB200 chips over 38 days.
