Yes, but that’s no contradiction, I still think Yamaha‘s SA2 sounds are more based on samples than Roland SN-A sounds. Roland has never given any sample sizes for its SN-A collection, which again supports that modeling plays a big role. AFAIK, Yamaha does state WaveROM sizes for Tyros 5 and Genos, in which the SA2 voices are included. Another hint that it‘s more scripting than modeling...
About using small samples for physical modeling: I think it‘s more a theoretical debate, actually I assume in one way or another sample data always flows into modeling of any acoustic instrument. GEM also stated that the RP-X piano module didn‘t use samples at all, but I‘m convinced that the modeling data has to be somehow derived from samples, otherwise the developers would have no idea how a waveform would have to be shaped to create a piano-like sound. It‘s just the question is it more rigid sample chunks put together or a very flexible modeling in which the waveform curve reacts totally flexibly, which would be real modeling. At least that‘s how I imagine his technology...