one year on
Meta's LLaMA large language model leaks online via 4chan torrent
A week after Meta restricted LLaMA to approved researchers, an anonymous user posted a torrent of the model's weights on 4chan, making frontier-adjacent AI freely downloadable for the first time.
Meta’s large language model LLaMA, released only a week ago under strict case-by-case approval for academic researchers, government officials, and civil society, has leaked online. An anonymous user on 4chan posted a torrent file containing the full model weights, making the 65-billion-parameter system—claimed to rival OpenAI’s GPT-3 and Google’s PaLM—freely available to anyone with a BitTorrent client.
The leak marks the first time a major tech firm’s proprietary AI model has escaped into the wild. Unlike OpenAI, Google, and Microsoft, which keep their newest models behind APIs or consumer interfaces, Meta had attempted a middle path: share with select researchers to balance openness and responsibility. That approach has now been circumvented. The leak spread rapidly: GitHub repositories hosting download instructions appeared, though Meta quickly filed takedown requests. Hugging Face CEO Clement Delangue warned users not to upload the weights, noting Meta’s copyright infringement claim.
4chan members claim to be running LLaMa on their own machines, though the exact implications of the leak are not yet clear. Experts have warned this technology could be used to automate fake news, spam, phishing emails, disinformation, incitement, and other misuse. AI engineer Shawn Presser, who posted download instructions on GitHub, told The Register the benefits will outweigh harms ‘by at least tenfold.’ Meta’s spokesperson stood by the limited release strategy, saying ‘we believe the current release strategy allows us to balance responsibility and openness’ despite the circumvention.
The record
An AI engineer who wrote download instructions on GitHub told The Register he believes a researcher with authorized access leaked LLaMA, and argued releasing the model freely is better: 'I think the good will outweigh the bad, by at least tenfold. Probably closer to 100x.'
Hugging Face CEO posted a staff update from GitHub showing Meta had filed a takedown request against a user's LLaMA repository, characterizing it as unauthorized distribution; Delangue cautioned users against uploading the weights.
One year later — open only if you can handle spoilers
This leak is now viewed as the birth of the open-weights era. LLaMA's unauthorized release directly inspired the creation of Alpaca, Vicuna, and a wave of fine-tuned variants that reshaped the AI landscape. By mid-2023, the LLaMA leak had shifted the industry's center of gravity from proprietary APIs to open models, leading Meta to eventually release LLaMA 2 as fully open source later that year.