Exactly, benchmarks are just a guide. Same thing with vector models. Likely just a matter of how much your data overlaps with the fine-tuning dataset for a particular model.