What others said ^ But note thst Meta's LLaMA 33B and 65B were trained with more tokens (1.2 trillion?) than the 13B and 7B models (1 trillion). And subjectiv

What others said ^ But note thst Meta's LLaMA 33B and 65B were trained with more tokens (1.2 trillion?) than the 13B and 7B models (1 trillion). And subjectively, the larger parameter models do indeed feel "smarter" beyond what objective metrics would suggest.