LLaMA: How Meta Opened the Model Weights and Changed AI

On February 24, 2023, Meta AI released LLaMA, Large Language Model Meta AI, as a research release. Access required an application. The weights were sent to approved researchers. Within a week, the weights had leaked to 4chan via a torrent link. Within two weeks, the open-source community had produced Alpaca, a fine-tuned version of LLaMA-7B that reportedly matched GPT-3 and cost approximately $600 to train. The controlled research release became one of the most consequential open-source distributions in the history of software.

Meta's motivation was not altruism. Meta operates at a scale where per-token API pricing would be prohibitive across Facebook, Instagram, and WhatsApp. Open-source AI infrastructure reduces dependency on OpenAI and Google, creates a talent pipeline, and, by establishing open models as competitive alternatives, reduces the monopoly pricing power of proprietary providers. The open release was a strategic move by a company for which openness was cheaper than closedness. The downstream ecosystem, thriving open-source LLMs, fine-tuned for thousands of use cases, was a side effect of corporate self-interest that benefited the entire field.

The community response exceeded any prediction. Georgi Gerganov built llama.cpp, a pure C++ inference engine that ran LLaMA on a MacBook. LoRA (Low-Rank Adaptation), adapted to produce personalised models with minimal GPU time. Vicuna, Koala, WizardLM, and dozens of derivative models each fine-tuned LLaMA on curated instruction data. By summer 2023, a solo developer with a $2,000 GPU could produce a capable instruction-following model. The democratisation of LLMs had happened faster than any institution had prepared for.

LLaMA 2 arrived in July 2023 with an explicit commercial licence. Llama 3, released in April 2024, was competitive with GPT-4 on many benchmarks. The open-source ecosystem had not just survived but accelerated: Mistral 7B (September 2023) matched Llama 2 13B with half the parameters; Mixtral 8x7B (December 2023) used a mixture-of-experts architecture to match GPT-3.5 while activating only 12.9 billion parameters per token. The open-source models were no longer two years behind the frontier: they were months behind, and closing faster.

The regulatory implications remain unresolved. The EU AI Act includes provisions for open-source models, ultimately providing a lighter compliance path, a concession to the argument that open weights enable safety research and reduce centralisation. Opponents argue that open weights cannot be recalled if misuse is discovered after release. The tension between the safety argument for openness (distributed scrutiny catches more problems) and the safety argument for closedness (limiting capabilities limits misuse) is the central unresolved debate in AI governance. LLaMA did not start the debate. It made it unavoidable.