I’m not sure if I understood the application correctly, but if you want to “bound” the number of drawn fruits from each type because you know how many of each type is available, this is basically taking draws without replacement. The multinomial distribution assumes draws with replacement (like the binomial). The multivariate hypergeometric (https://en.m.wikipedia.org/wiki/Hypergeometric_distribution#Multivariate_hypergeometric_distribution) is maybe the thing you are looking for. Like the hypergeometric districtribution, it assumes drawing without replacement. Parameters are the number of available items to draw from, the number available for each type, and the number of items drawn. Those are all integers, so if any of those are unknown or only partly observed in the data, they would need to be integrated over in Stan.