LLaMA

LLaMA (Large Language Model Meta AI) is Meta's first foundation-model family, released February 24, 2023 in four sizes: 7B, 13B, 33B, and 65B parameters. Trained on 1.4 trillion tokens of publicly available text, it showed that smaller models trained on more data could match or beat much larger ones — the 13B variant outperformed the 175B GPT-3 on most benchmarks.

Meta released it under a non-commercial research-only license, with weights available to approved researchers on request. Within days of release, the weights leaked publicly, which — combined with Meta's permissive-by-accident posture — ignited the explosion of open-weight fine-tunes (Alpaca, Vicuna, and hundreds more) that defined 2023.

An honest caveat: LLaMA was never truly open-source. The original license was research-only and non-commercial, and the leaked weights existed in a legal gray area. Its successors (Llama 2 and 3) adopted more permissive community licenses, but the original LLaMA remains a research artifact of historical importance.