1) to use a branch model:

• if you want to let w free to have different values in each branch, you have to set the variable model = 1 (this is not usually a good choice because you may overfit the data and have meaningless results).
• you can group some branches together and associate them with the same w which can either be fixed or not. in this case you should set model = 2. by doing so you lose resolution but the results you get are more likely to be relevant (to know how to group branches you can have a look at the example/lysozyme/)

2) to use a site model set the variable model = 0 and use the variable Nssites to choose the models you want to evaluate (e.g. Nssites = 1 2). you can then compare two models by means of the likelihood ratio test, to see which one can best explain the observed data.

these analyses require codon type sequences (seqtype = 1). all the concepts and statistics you can find in paml are widely discussed in the book "computational molecular evolution". I would also suggest you to formulate an hypothesis about your data, before running many different analysis. if you don't know what you are looking for, you may find anything by chance!

try its examples first.