A mixture-of-experts model contains many sub-networks, called experts, together with a small router that decides which of them should handle each piece of input. For each token, only a few experts are used and the rest stay idle. The experts are not experts in a human sense, and they are not necessarily assigned to neat topics. They specialize as a result of training.
The advantage is efficiency. The model can hold a very large number of parameters, which adds capacity, while the amount of computation for each token stays closer to that of a much smaller model. That can make it faster and cheaper to run than a dense model of the same total size.
The costs are memory and complexity. All of the experts must be stored and available, even if only some are used at a time, so memory requirements follow the total size. Training can be harder, because the router has to balance the load between experts, and serving requires careful engineering.
Compare total and active parameters, and judge the model on speed, cost and quality for your use.