Router

A router is like an information desk inside a model. Think of the whole model as a large hospital, with different departments handling different jobs — internal medicine, surgery, pediatrics. Each of these departments is called an expert. Every time a single patient, meaning a single input, comes in, the router decides on the spot which department to send it to.

Models built this way can have a huge overall size, yet only a fraction of the departments actually work at any given moment. The hospital may employ a large staff overall, but when treating one patient, only a few examination rooms actually light up. Depending on how many experts the router calls up for each input, and how often, the actual computing cost at that moment changes.

The catch is that the exact criteria the router uses to pick experts, and exactly how many it picks, are often known only inside the company that built the model. Companies proudly announce the total size (total parameters) in their releases, but frequently keep the number that actually does the work each time (active parameters) undisclosed.

In the article, Qwen3.8-Max itself described active parameters as "the scale of parameters actually used in computation, through the subset of experts selected by the router when processing a specific input token." A common misunderstanding: the router doesn't pick departments at random — it judges based on what it learned during training about which experts suit which inputs.

Try asking an AI chatbot known to use a mixture-of-experts architecture: "Are you a mixture-of-experts model? If so, do you know how many experts the router picks for a single response?" Most will say they don't know the exact number, or that the developer hasn't disclosed it. That answer itself is a good illustration of this concept.