Active Parameters

An AI model consists of billions of numbers that were tuned during training. These numbers are called parameters, and they store the model’s entire knowledge. In older models, every response involved computation with all of these numbers. Newer models do it differently: they switch on only a portion of their numbers per request. Exactly this switched-on portion is the Active Parameters. A model can thus possess 400 billion numbers and still answer using only 20 billion of them.

Why the total count alone says nothing about cost

For a long time, a simple rule held: more parameters means more computation and higher costs. With active parameters, this rule falls apart. The compute time for a response depends almost entirely on how many numbers participate in the calculation. How many exist in total plays hardly any role in that.

For companies, this is an enormous lever. Anyone running a chatbot with millions of users pays for electricity and compute time for every single response. If the number of active parameters drops to a tenth, these costs fall drastically. The model still remains large and thus versatile, because the knowledge in the unused parts doesn’t simply vanish.

That’s why providers today usually state two numbers. A figure like “235B total, 22B active” means: 235 billion parameters exist, 22 billion compute per request. Anyone reading only the first number massively overestimates the operating costs. Anyone reading only the second underestimates the memory requirements.

How the switching on and off comes about

This is made possible by an architecture called Mixture of Experts. The model is divided into many sub-networks called experts. A small additional network, the router, decides for each word which two or four experts are responsible. All others remain silent for that request.

A comparison helps: a library with a million books is huge, but for a homework assignment you pull three books off the shelf. The remaining 999,997 cost you no reading time. But they still have to stand in the building, otherwise you could never fetch them.

This also names the most important misconception. Active parameters save compute time, not storage space. The complete model must reside in the GPUs' memory, including the dormant parts. A model with 400 billion parameters therefore still needs very expensive hardware, even if only 17 billion of them are active. Large models become faster this way, not smaller.

Where this figure shows up in product announcements

In the model cards and blog posts of the major providers, this figure is now given almost always. Mistral, DeepSeek, Alibaba's Qwen series, and Meta's Llama models list active and total parameters separately. Trade media and analyst reports also use it as an argument when discussing cost per response.

In practice, you notice the effect in the pricing of programming interfaces. A model with few active parameters often costs only a fraction of an equally large older model per unit of processed text. Response speed in chat depends on it too.

Caution is warranted with comparison tables. If a provider markets its model as “smaller and more efficient,” it’s worth checking which of the two figures is meant. A model with 3 billion active parameters can in reality be 30 billion in size, and thus won’t fit on a normal laptop.