In natural sciences, we usually consider the system A to be a model of the system B if A is similar to B in some important properties and exhibits somewhat similar behavior in similar circumstances. According to this definition, it is unnecessary the physical nature of the modeling system A be the same as of the modeled system B. For example, B can be a technological aggregation in chemistry, while A can be a set of differential equations, i.e., a conceptual system of a mathematical nature. A mathematical model is usually the best if it really ensures sufficient approximation.
In the linguistic domain, we need a modeling system that, after receiving information of a linguistic nature in its input, exhibits in its output an acceptable degree of similarity to the results of natural language activity of the human brain. Such a definition follows the idea of Alan Turing mentioned above.