Active Parameters
Active parameters are the portion of an AI model's adjustable numerical values that actually participate in the computation for a single request. For some models, this is only a fraction of the total, which makes operation significantly cheaper.
An AI model consists of many billions of numbers that are adjusted while learning from examples. These numbers are called parameters, and they contain everything the model is capable of. Not every model uses all of these numbers for every request. Some models only switch on a portion of them per request. Exactly this switched-on portion is the active parameters. So a model can have 400 billion parameters and still only use 20 billion of them when answering a question.
Why this matters
The computation time for a response depends mainly on the active parameters. The more numbers that participate in the computation, the longer it takes and the more electricity it costs. The total number of parameters, on the other hand, tends to indicate how much knowledge is stored in the model.
That’s why this figure is doubly interesting for companies. They can advertise a large total count while still offering a cheap service. So anyone reading headlines about model sizes should always check which of the two figures is meant. A model with few active parameters runs faster, but still requires a great deal of working memory.
How does it work?
The common approach is called Mixture of Experts. In this setup, the model is broken down into many small subnetworks called experts. A small additional network, the router, selects a few of them for each word. Only these selected experts do the computing, while the rest remain idle at that moment.
A comparison helps: a large library has millions of books. For a particular question, you take three of them off the shelf. The remaining books still have to be in the building, otherwise you could never retrieve them. In the same way, the complete model has to sit in the memory of the graphics cards, even though only a fraction of it is actually working.
This leads to an important rule of thumb. Active parameters determine the speed and the cost per response. The total number determines how expensive the hardware is on which the model can run at all. Only both figures together fully describe a model.
Where you encounter the term
In technical announcements of new models, this figure is almost always given. Typical phrasings are things like “671 billion parameters, of which 37 billion are active”. Such figures appeared, for example, with models from DeepSeek, Mistral, and Alibaba. Several models from Google and Meta now also work according to this principle.
For investors, the figure is an indicator of operating costs. A provider with few active parameters can be cheaper per response than the competition. This explains why some Chinese models were able to offer such low prices in 2024 and 2025.
In everyday use, you don’t notice the difference directly. You just see that a chatbot answers quickly or is usable for free. Behind this often lies the decision of how many parameters are actually allowed to participate in the computation per request.