Activated Parameters
Activated parameters are the portion of an AI model's learned settings that actually factor into computation for a single request. This number reveals more about speed and cost than the often-cited total size of the model.
An AI model consists of many billions of numbers that were adjusted during training. These numbers are called parameters. They represent the accumulated knowledge of the model, comparable to the positions of countless tiny sliders. In some models, however, not all of these numbers are used for a single question. Only a portion is switched on and takes part in the computation. This exact portion that gets used is called activated parameters. The rest remains silent for this particular request, even though it still exists.
Why total size alone is misleading
Reports about new AI systems usually mention a large number: 400 billion parameters, 600 billion, sometimes more. This number sounds impressive but says nothing about speed and cost. If only 30 out of 600 billion numbers factor into each request, the model works about as fast as a much smaller one. That’s why manufacturers today often provide both figures: the total count and the activated count.
The activated count determines how much computational work each individual response requires. And every response costs electricity and computing time, repeated for millions of users every day. A model with few activated parameters is significantly cheaper to run. For companies offering such systems, this is a central cost factor.
A common misconception: many people read a small activated count as a sign of a weak model. That’s not accurate. The knowledge resides in the total amount, and for each question the appropriate slice of it is retrieved. Such a model can know more than a small model with the same computational speed.
How the selection within the model comes about
This is made possible by an architecture called Mixture of Experts. The model is divided into many subgroups, called experts. A small additional network, the router, decides for each word which two or three experts should work. Only their parameters count as activated. All other experts are simply skipped in the computation.
However, part of the model always participates in the computation, such as the building blocks that break down text and establish the relationships between words. The activated count is therefore the sum of this fixed portion and the currently selected experts. This method saves computing time, but not storage space: the dormant experts also have to reside on the graphics cards. Otherwise they would first need to be loaded, which would be slower than any savings gained.
This should not be confused with quantization. There, the same parameters are stored more coarsely to save space. With activated parameters, precision stays the same; only fewer of them are used simultaneously. Both techniques can be combined.
Where the number appears in spec sheets
Anyone looking at model cards or announcements often finds figures like “235B total, 22B active”. The B stands for billion, total for the total count, active for the activated parameters. Models from DeepSeek, Mistral, Alibaba, and Meta have been described this way for several years now. The large commercial systems are also likely built this way, but they do not publish their figures.
The difference becomes practically relevant when someone wants to run a model themselves. The total count determines how much graphics card memory is needed. The activated count determines how many words per second come out. So a model can simultaneously be too large for one’s own machine and pleasantly fast.
Activated parameters appear indirectly in stock market reports. If they decrease, the operating cost per response drops, and a provider can lower prices or increase profit margins. That’s why a low activated count combined with good quality is considered a technical advantage that analysts write about.