Meta's LLaMA — A Research Release, a Leak, and the Open-Weights Debate

LLaMA (Large Language Model Meta AI) is the large language model Meta announced on 24 February 2023. It came in four sizes, 7, 13, 33 and 65 billion parameters, released under a non-commercial research licence to researchers and others approved case by case. About a week after access began, on 3 March, the model's weights leaked on 4chan and anyone could run it on their own hardware. The leak, and Llama 2, the successor Meta released for commercial use on 18 July, put model weights at the centre of the open-versus-closed debate in AI.

It came less than three months after the launch of ChatGPT, and it marked the moment when powerful language models stopped being something that existed only behind the APIs of a few companies such as OpenAI and Google.

The February 2023 announcement

Meta presented LLaMA as a foundation model to help researchers, arguing that smaller, more performant models would let people without large amounts of infrastructure study these systems, "further democratizing access in this important, fast-changing field."

The paper, "LLaMA: Open and Efficient Foundation Language Models", claimed to show that state-of-the-art models could be trained on publicly available data alone, with no proprietary datasets, and that the 13-billion-parameter model outperformed GPT-3 (175 billion) on most benchmarks while the 65-billion model was competitive with Chinchilla-70B and PaLM-540B. Those are the Meta researchers' own assessments. The focus on Latin and Cyrillic languages also meant Japanese would need extra work.

The leak on 3 March

According to The Verge, just one week after Meta started fielding access requests, on 3 March, a downloadable torrent was posted on 4chan and spread across AI communities. Researchers who had downloaded it told the magazine it was legitimate.

Reactions split. "Get ready for loads of personalized spam and phishing attempts," tweeted cybersecurity researcher Jeffrey Ladish. "Open sourcing these models was a terrible idea." Sayash Kapoor and Arvind Narayanan countered that a wave of malicious use of language models had long been predicted, "Yet, there don't seem to be any documented cases." It was a contest between those who blamed Meta for distributing the technology too freely and those who argued that open access is needed to build safeguards.

The leaked weights quickly became a research platform. On 13 March Stanford researchers introduced Alpaca, LLaMA 7B fine-tuned on 52,000 instruction-following demonstrations, which they said behaved qualitatively similarly to OpenAI's text-davinci-003 on single-turn instructions while being reproducible for under $600. The public demo was later taken down, citing hosting costs and the inadequacies of its content filters.

Politics followed. According to Decrypt, on 6 June Senators Richard Blumenthal, chair of the Senate's Subcommittee on Privacy, Technology and the Law, and Josh Hawley, its ranking member, wrote to Zuckerberg citing the "potential for its misuse in spam, fraud, malware, privacy violations, harassment, and other wrongdoing and harms," and asked how Meta had assessed the risks before release.

Llama 2 on 18 July, and the meaning of "open source"

On 18 July 2023 Meta released Llama 2, with Microsoft as its preferred partner. The weights of the pretrained and chat-tuned models, at 7, 13 and 70 billion parameters, were free for research and commercial use and available through Azure, AWS, Hugging Face and others. Meta said Llama 1 had drawn more than 100,000 requests for access from researchers.

Meta called Llama 2 "open source", but section 2 of its licence says that if, on the release date, a licensee's products had more than 700 million monthly active users in the preceding calendar month, it must request a licence that Meta may grant at its sole discretion. On 20 July the Open Source Initiative, which maintains the Open Source Definition, said the licence fails that definition because it restricts commercial use for some users and, through its Acceptable Use Policy, certain purposes, while open source may not discriminate against persons, groups or fields of endeavour. Meta, the OSI wrote, was confusing "open source" with "resources available to some users under some conditions." Models whose weights alone are published have since increasingly been called "open weights".

The ripple in Japan

Because commercial use was allowed, Japanese-language models were built on Llama 2. On 29 August 2023 ELYZA released ELYZA-japanese-Llama-2-7b and related 7-billion-parameter models, created by further pre-training Llama 2 on Japanese text, usable for research and commercial purposes as long as users follow the Llama 2 licence and Acceptable Use Policy. ELYZA described their performance as comparable to GPT-3.5 (text-davinci-003) and the best among publicly available Japanese models. For companies and labs without the money to train a huge model from scratch, published weights offered a starting point.

What it meant, and the swing back

LLaMA raised the question of who should hold capable models, who gets to inspect them and who is responsible for them. In his letter "Open Source AI is the Path Forward" on 23 July 2024, Zuckerberg argued that AI would follow Linux, which beat the closed Unixes, and released the 405-billion-parameter Llama 3.1. In January 2025 China's DeepSeek built some of the small models distilled from DeepSeek-R1 on Llama.

Meta's own stance then shifted. In a 30 July 2025 essay Zuckerberg wrote that superintelligence would raise novel safety concerns and that Meta would need to be "careful about what we choose to open source." According to Fortune, on 8 April 2026 Meta unveiled Muse Spark, the first model from its new Meta Superintelligence Labs, without releasing the weights, offering it only in Meta's own apps and as a private-preview API for selected partners. Fortune also noted that Llama 4, released in April 2025, had been widely panned, and that Meta admitted using specially tuned, unreleased versions to boost its benchmark scores. Meta says it hopes to open-source future versions, but the open-weights path that began in 2023 is now being questioned even inside the company that led it.

Other AI events are on the AI timeline and the open-source timeline, and the company's history on the Meta page.