Hardware Ready Ready to Bench?
Microsoft MAI Base 2. Microsoft has presented MAI Base 2, a new artificial intelligence model based on a Mixture of Experts (MoE) architecture that employs a total of 2.8 trillion parameters, with a portion of 80 billion of these active during execution. The company highlights how this approach can lead to significant advantages in terms of computational efficiency.
According to the Microsoft AI division, the new system would guarantee pre-training efficiency that is 4 to 12 times higher, measured in FLOPs (Floating Point Operations Per Second), compared to DeepSeek V4 Pro Base. This statement positions Microsoft decisively in the competition for the development of next-generation foundational models.
Microsoft MAI Base 2: why it matters
The term “Base” in the model’s name indicates that it is the pre-trained version, meaning the fundamental structure upon which versions optimized for specific tasks will be built. The reference to DeepSeek also underscores this aspect: Microsoft chose to compare two systems at the same stage of development.
How a 2.8-trillion-parameter MoE model works
To understand the significance of these figures, it is useful to explain how a Mixture of Experts (MoE) model works. In a traditional architecture, every processed request traverses the entire neural network, involving all its parameters. With the MoE approach, however, the model is divided into numerous specialized blocks called “experts,” and only a portion of them is activated for each operation.
The result is a system that is imposing in terms of overall size but operates with a fraction of its resources at any given time. In the case of MAI Base 2, out of a total of 2.8 trillion parameters, only 80 billion are active, which is less than 3%. This configuration allows for maintaining broad overall capacity without incurring the computational cost derived from using the entire structure at every step. It is precisely in this resource management that a significant part of the potential efficiency gains resides.
The efficiency challenge against DeepSeek V4 Pro Base
The core of the announcement lies in the direct comparison with DeepSeek. Microsoft claims that MAI Base 2 is 4 to 12 times more efficient in pre-training than DeepSeek V4 Pro Base, using FLOPs as the reference metric. FLOPs represent the number of floating-point operations executed by processors and are a common indicator of the computing power required to train an artificial intelligence model.
Microsoft specifies that the range between 4x and 12x refers to efficiency during pre-training, which is the initial phase in which the model learns from large-scale data. It should be noted that this is a claim by the company itself, presented by Microsoft AI alongside the new model. The comparison with DeepSeek V4 Pro Base concerns exclusively the amount of computation employed in this preliminary phase and is based on systems in their base version, before any specific optimizations.
Microsoft did not provide a single value for the efficiency observed but rather a variation between 4 times and 12 times. This difference could depend on various factors related to the hardware and software configurations used during testing.
Source and further reading on Microsoft MAI Base 2: original article.
* Content created with the assistance of artificial intelligence systems.