Microsoft has completed pre-training its latest artificial intelligence model, MAI-Base-2, featuring an impressive 2.8 trillion total parameters. The technology giant claims the new system outperformed every open-weight AI model included in its internal evaluations while delivering substantial improvements in training efficiency.
The announcement represents another major milestone in Microsoft’s efforts to develop advanced AI technology independently and strengthen its position in the increasingly competitive artificial intelligence industry.
According to the company, MAI-Base-2 could deliver comparable results to a leading rival using significantly fewer computing resources.
Microsoft MAI-Base-2 Features 2.8 Trillion AI Parameters
Microsoft’s new AI model uses a mixture-of-experts (MoE) architecture, allowing it to activate only a fraction of its total parameters when processing information.
Although MAI-Base-2 contains 2.8 trillion parameters, approximately 80 billion are active during each processing task.
This design allows the system to benefit from an exceptionally large model architecture without requiring every parameter to operate simultaneously.
As a result, Microsoft aims to improve computational efficiency while maintaining strong performance across complex AI workloads.
The company said MAI-Base-2 surpassed all open-weight models evaluated in its internal pre-training benchmarks.
However, these results have not yet been independently verified, and the model is still undergoing additional development.
Microsoft Claims Up to 12 Times Greater AI Training Efficiency
One of the most significant aspects of MAI-Base-2 is its reported improvement in computing efficiency.
Microsoft compared its training approach against DeepSeek-V4-Pro-Base, a competing open-weight foundation model.
Across four internal evaluations, the company estimated that its training methodology was between 4.2 and 12.3 times more compute-efficient.
The comparison focused on the processing power required to achieve equivalent benchmark results during pre-training.
These figures do not necessarily indicate that the finished model will generate responses 12 times faster or operate at one-twelfth of the inference cost.
Nevertheless, the findings suggest that Microsoft has made substantial progress in optimizing the development of large-scale AI systems.
The company attributed these improvements to better training data, architectural refinements and more efficient computing infrastructure.
Nvidia GB200 GPUs Power Microsoft’s Massive AI Training Project
Developing MAI-Base-2 required significant computing resources.
Microsoft trained the model using approximately 16,000 Nvidia GB200 GPUs over a 38-day period spanning August and September 2026.
The training process involved roughly 50 trillion tokens, reflecting the enormous amount of information required to develop advanced foundation models.
Compared with MAI-Base-1, which Microsoft introduced in June, the new model benefited from a training dataset containing approximately 36% more unique tokens.
Microsoft also reported substantial improvements in its training methodology compared with the previous generation.
According to its internal calculations, the MAI-Base-2 training approach achieved efficiency gains ranging from 6.3 to 12.7 times those of MAI-Base-1.
These advances highlight the company’s focus on developing more capable artificial intelligence systems while optimizing the computational resources required to train them.
MAI-Base-2 Expands Into Vision and Multimodal AI
Another important development is Microsoft’s decision to incorporate visual information directly into the model’s training process.
Unlike its predecessor, MAI-Base-2 was designed with native vision capabilities, allowing it to learn from both text and image-related information.
Visual training material accounted for approximately 18% of the model’s overall training mixture.
This approach could help the system perform tasks involving image interpretation, visual reasoning and other applications that require understanding information beyond traditional text.
Microsoft believes these capabilities will become increasingly important as businesses adopt more sophisticated AI tools.
By combining large-scale language processing with visual understanding, the company aims to create foundation models capable of supporting a broader range of enterprise and consumer applications.
Microsoft Focuses on Independent AI Development and Azure Integration
Beyond improvements in model performance, MAI-Base-2 represents an important step in Microsoft’s broader artificial intelligence strategy.
The company is investing heavily in proprietary AI development to gain greater control over the technology supporting its products and cloud services.
Microsoft emphasized that MAI-Base-2 was trained using original data rather than relying on knowledge distilled from competing AI systems.
According to the company, this distinction provides greater control over the provenance of its training information.
It could also support Microsoft’s plans to license its AI model weights to selected enterprise partners through Azure AI Foundry.
Such an approach would allow Microsoft to expand its artificial intelligence offerings while reducing its dependence on external model developers.
Greater ownership of its AI infrastructure could also provide more flexibility in product development, commercial licensing and long-term technology investment.
Microsoft Already Preparing MAI-Base-3 as AI Competition Intensifies
Although MAI-Base-2 has completed its initial pre-training phase, the model is not yet a finished commercial product.
Microsoft confirmed that post-training remains underway, with additional performance results expected as development progresses.
The company has not announced an official commercial release date or disclosed pricing details.
Meanwhile, preparations for the next-generation MAI-Base-3 model have already begun.
This suggests Microsoft intends to maintain an aggressive development schedule as competition intensifies among leading artificial intelligence companies.
The latest announcement also highlights a broader industry trend toward increasingly sophisticated AI architectures that combine massive parameter counts with more efficient computing methods.
For Microsoft, achieving greater training efficiency could prove particularly important as demand for advanced AI systems continues to grow.
However, the model’s ultimate competitive position will depend on its performance after post-training, independent evaluations and eventual commercial deployment.
For now, MAI-Base-2 demonstrates Microsoft’s ambition to become a leading developer of frontier AI models while strengthening its technological independence.
