Meta released LLaMA in four sizes — 7, 13, 33, and 65 billion parameters — distributing the weights only to researchers and government applicants who passed an approval process, under a non-commercial research licence. The paper reported that the 13B model beat GPT-3 (175B) on most benchmarks and that the 65B was competitive with DeepMind's Chinchilla-70B and Google's PaLM-540B: train a smaller model for longer, and it catches up with far larger ones. Handing out weights instead of opening an API was a limited, research-only measure at this point, but it put a frontier-class model that ran on modest hardware outside the companies that built it.