A request form, not a download button
Meta released LLaMA in four sizes, 7B, 13B, 33B and 65B parameters, but did not put the weights up for anyone to take. Access went through an application, granted case by case to academic researchers and to people affiliated with government, civil society and academia, under a noncommercial research license. [1]
Meta described the gate as a balance rather than a restriction: releasing a capable model at all, while keeping the recipient list to people it could vouch for. [1]
The torrent that beat the approval queue
On March 3, 2023, a week after the announcement, a 4chan user posted a magnet link to LLaMA's full weights. Within days a pull request appeared against Meta's own repository asking to add that link to the documentation, and copies spread to GitHub and Hugging Face faster than Meta could file takedowns. [2][3]
Meta did not deny the leak. Vice quoted a spokesperson saying that while the model was not accessible to all, and some had tried to circumvent the approval process, the company still believed its release strategy balanced responsibility and openness. [2]
What showed up before Meta ever answered a request
Within days, engineers had squeezed the leaked weights onto hardware they were never sized for: the 65B model on a single Nvidia A100, the 13B model on a MacBook Pro. A 4chan-built Discord bot nicknamed BasedGPT used the leak to generate content the license would have barred, while Stanford's Alpaca team, working from weights obtained through the approved channel, held its own fine-tuned model back rather than publish over Meta's objections. [3]
Atlas interpretation: The split matters more than the leak itself. The researchers who respected the license waited for permission; the people who didn't already had the model running. A distribution plan that depends on everyone downstream behaving only works until the first person who doesn't, and LLaMA's gate lasted exactly the week it took for that to happen. [2][3]
Five months later, Meta stopped gating the same model line
Meta released Llama 2 on July 18, 2023, this time with no application form: free for research and commercial use, with model weights and starting code posted directly and distributed through Azure, AWS and Hugging Face. The announcement noted more than 100,000 requests had come in for access to the original LLaMA, framing the wide release as scaling up demand rather than reversing a policy. [4]
Atlas interpretation: Meta never credited the leak with the change, and a hundred thousand pending requests is itself a case for opening the gate on its own. But the sequence is hard to miss. The company spent five months watching a leaked, license-violating copy of its model become the de facto open release, then shipped the next version under terms close to what the leak had already established in practice. [1][4]
Sources
- Introducing LLaMA: A foundational, 65-billion-parameter large language model
Meta AI · Feb 24, 2023
- Facebook's Powerful Large Language Model Leaks Online
Vice · Mar 7, 2023
- Runaway LLaMA: How Meta's LLaMA NLP Model Leaked
DeepLearning.AI · Mar 15, 2023
- Meta and Microsoft Introduce the Next Generation of Llama
Meta AI · Jul 18, 2023