The History of AI

LLaMA

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models

Lineage
Neurons and networks
Kind
Implementation

Who

Meta

What happened

Meta presents LLaMA, a family of language models with 7 billion to 65 billion parameters trained only on public data, and makes it available to researchers under a noncommercial license. A week later its weights leak onto the internet.

Why it mattered

The open-weight model ecosystem takes off: a competitive model can be run and fine-tuned outside the big labs.

The full story

In early 2023 the most capable language models were used through their owners' services. On February 24 Meta presented LLaMA, a family of four models with 7, 13, 33, and 65 billion parameters. The paper, by Hugo Touvron, Guillaume Lample, and twelve other researchers, appeared on arXiv three days later.

The bet was on efficiency. The models were trained only on public data, and according to the paper the 13-billion-parameter model beat GPT-3, which had 175 billion, on most benchmarks, while the largest was competitive with Chinchilla and PaLM, two of the best models of the time. Meta argued that small models, since they need far less compute, opened research up to those without large infrastructure.

Meta did not simply publish the weights: it granted them case by case to researchers at universities, government bodies, civil society organizations, and industry labs, under a noncommercial license. On March 3, a week after the announcement, they were spread over BitTorrent by way of the 4chan forum.

Derivative models were built on those weights, such as Stanford University's Alpaca, based on the smallest LLaMA. With Llama 2, in July 2023, Meta allowed commercial use, and it followed with Llama 3 in April 2024 and Llama 4 in April 2025.

Common misconceptions

  • What people often say: Meta released LLaMA openly for anyone to use.

    In fact: Meta granted the weights case by case to researchers, under a noncommercial license. They circulated freely because they leaked a week later, and commercial use was not allowed until Llama 2.

Builds on

  • Chinchilla (March 29, 2022)

    Documented direct influence

    LLaMA starts from the scaling laws of Hoffmann et al. (2022) and chooses to train smaller models on more tokens than those laws recommend, with the cost of running them in mind (section 1)

Sources

How to cite

“LLaMA.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/llama

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline