Microsoft Unveils MAI-Base-2, a 2 Trillion Parameter Model Built for Efficiency
Microsoft's new base model outperforms DeepSeek's leading open-source alternative on key benchmarks while consuming significantly less computational power, thanks to an all-mixture-of-experts architecture.
Microsoft has unveiled MAI-Base-2, a foundational AI model designed to deliver superior performance while demanding far fewer computational resources than competing open-source offerings. The artificial intelligence research division at Microsoft announced the breakthrough on Thursday.
Unlike a finished large language model, MAI-Base-2 functions as a base model—essentially a raw foundation that developers refine through additional training stages. These subsequent training phases typically introduce reasoning capabilities and safety mechanisms before the model reaches production use.
When tested against DeepSeek-V4-Pro-Base, the most sophisticated open-source base model currently available, MAI-Base-2 demonstrated superior results on Microsoft's internal evaluation suite covering science, programming and mathematics. The new model also achieves up to 12 times greater hardware efficiency by certain metrics.
Architecture and Design
MAI-Base-2 employs a mixture of experts, or MoE, architecture that activates only a fraction of its neural network when processing requests. Rather than engaging the complete network, this selective activation approach reduces computational overhead. The model contains 2 trillion total parameters but only engages 80 billion of them per prompt.
Microsoft's earlier base model combined MoE layers with FFN modules, a less efficient design approach. MAI-Base-2 abandons this hybrid strategy in favor of an all-MoE composition throughout its entire structure, contributing significantly to its performance gains.
Training and Development
The model underwent 38 days of training using 16,000 of Nvidia's GB200 chips. Microsoft's training dataset encompassed 50 trillion tokens, with approximately one-fifth dedicated to developing an integrated computer vision component. This vision capability enhances MAI-Base-2's ability to interpret scientific diagrams and similar visual content compared to its predecessor.
During the training process, Microsoft implemented gradient clipping to mitigate issues that can emerge when mathematical gradients—the feedback signals that refine model performance—propagate through network layers. The company also leveraged Muon, an established open-source optimization tool, to accelerate the training workflow.
Deployment and Future Plans
Microsoft intends to make MAI-Base-2's model weights available to "trusted partners" via its Azure cloud infrastructure. The company will also integrate the base model into a forthcoming lineup of language models, with additional details expected to emerge in coming weeks.
Read more
Nvidia shifts from selling chips to selling AI infrastructure as a complete system
At GTC Berlin, Jensen Huang repositioned Nvidia's strategy away from graphics processors toward end-to-end AI deployment, introducing containerized factories, agent security frameworks, and structured data models that address enterprise bottlenecks beyond raw computing power.
Five Technology Shifts Will Push CIOs From Passive Oversight to Active Design
Enterprise leaders must architect AI systems proactively rather than react to rapid technological change, according to Gartner research presented at its IT Symposium/Xpo in Orlando.
Comet Cloud eyes underserved SMBs as GPU inference demand shifts
With AI workloads moving from training to inference, neocloud providers see opportunity in small and midsize businesses that hyperscalers have deprioritized. Comet Cloud is positioning itself to serve that market with dedicated clusters and aggressive pricing.