The Original LLaMA Release and What It Set Off

Meta's first LLaMA, announced in February 2023, was a family of foundation models at 7B, 13B, 33B, and 65B parameters. The larger models were trained on 1.4 trillion tokens, the 7B on one trillion. Meta released them under a noncommercial research-only license, with access gated by application for academics, government and civil society groups, and industry labs. The framing was about making large language model research possible for groups without huge infrastructure. The accompanying research paper, not the blog post itself, is where the well-known result lives that LLaMA-13B performs competitively with the much larger GPT-3 175B, so attribute that to the paper rather than the announcement. What actually changed the field was not the official program. The weights leaked, and within weeks a wave of projects built on them: Alpaca, Vicuna, and llama.cpp, which made the models run on ordinary hardware.

Why it matters

This is the release that created the modern open-weight ecosystem, mostly through a leak Meta did not plan. If you use any locally runnable open model today, the tooling and the expectation that such models exist trace back to this moment. It is also a case study in how little control a lab keeps once weights are out.