Meta releases LLaMA to researchers, and it leaks within a week

Meta released LLaMA, a family of foundation models from 7 to 65 billion parameters, to approved academic researchers under a non-commercial licence. Trained on publicly available data and following Chinchilla-style token budgets, the 13-billion-parameter version was reported to outperform GPT-3 on most benchmarks despite being more than ten times smaller.

A week later the weights were posted to 4chan and spread via BitTorrent. Meta did not pursue takedowns aggressively, and the leaked models became the substrate for a fast-moving open ecosystem. Stanford’s Alpaca, released in March, fine-tuned the 7B model on 52,000 instruction-following examples generated by GPT-3.5 for a few hundred dollars in compute. Vicuna followed, then llama.cpp, which made the models run on laptops and eventually phones via quantisation. A leaked internal Google memo in May, titled “We Have No Moat,” argued that this community was iterating faster than either Google or OpenAI.

The episode shaped Meta’s strategy. Llama 2, released that July with weights available for commercial use, formalised what the leak had made true, and Meta spent the following years as the principal industrial advocate for open weights — a position that put it in direct conflict with proposed regulation premised on controlling model distribution.

It also established the terms of a debate that had no clean resolution: whether publishing weights mainly democratises capability and enables safety research, or mainly removes the ability to withdraw a model once its harms become apparent. Both claims were true, and the leak meant the experiment ran regardless.