Active Parameters

Active parameters refer to the size of the part of a model that actually gets switched on to process a given question. If the whole model is a huge library, the total parameters are every book on every shelf, while the active parameters are the handful of books actually pulled down to answer one question. The library can be built as large as you like, but since you don't need to search every shelf each time — just open the relevant sections — you gain in both speed and cost.

Models built this way usually list both the total parameter count and the active parameter count together. For example, '30 billion total, 3 billion active' means the model's overall stored knowledge is on the scale of 30 billion parameters, but the computation and response speed for handling any single question is closer to that of a 3-billion-parameter model. The core appeal of this structure is stacking up broad knowledge while keeping the per-use computational load light.

However, some companies only publicize a large total parameter figure without disclosing the active parameter count. In that case, it becomes hard to estimate how much it actually costs to run the model.

One article explained that a model has 30 billion total parameters but only 3 billion actually participate in computation when generating a single token, noting this structure makes it faster than similarly sized models. In contrast, another article covered how Alibaba released a 2.4-trillion-parameter model without ever disclosing its active parameter count, leading foreign outlets to estimate it differently — some at 95 billion, others at 22 billion. A common misunderstanding is judging a model's real operating cost from the total parameter figure alone; to gauge actual computational cost, you need to look at the active parameter number instead.

Try asking a chatbot the following to see how the concept of active parameters plays out in an actual answer:

"Tell me the total number of parameters in the model you're using, and how many parameters are actually activated when processing a single question. If you're not sure, just say you don't know."

It's rare for a model to give exact figures. But if you get an answer explaining why the two numbers differ, and what total versus active parameters each mean, that itself confirms you've grasped the concept.