Is there any rule of thumb for small language models vs large language models? I've seen phi 4 called a small language model but with 14 billion parameters, it's larger than some large language models.