The Llama leak refers to the unauthorized distribution of Meta AI's LLaMA (Large Language Model Meta AI) model weights in March 2023. This event marked a turning point in the development of large language models, as it bypassed Meta's controlled access policy and made the weights available to the public via BitTorrent and HuggingFace. The leak accelerated open-source AI research by enabling a broader community of developers and researchers to experiment with and build upon the models, despite initial concerns about misuse.
The leak occurred shortly after Meta announced LLaMA on February 24, 2023, as a family of foundation models ranging from 1 billion to 65 billion parameters. Unlike contemporaneous models like OpenAI's GPT-3, which were closed, LLaMA's weights were intended for academic researchers under a non-commercial license. The leak democratized access, leading to a proliferation of fine-tuned variants and tools, and influenced Meta's subsequent releases of Llama 2 and Llama 3, which were made more openly available.
Background
In the early 2020s, the field of artificial intelligence experienced a surge in interest following the release of OpenAI's ChatGPT in November 2022. This sparked a race among tech companies to develop more powerful large language models. Meta, under its AI research division Meta AI, pursued a strategy of open research, but with restrictions on model access. Meta's Chief AI Scientist Yann LeCun expressed skepticism about the hype, stating that large language models are best for aiding with writing, not for achieving human-level intelligence.
The LLaMA project aimed to create efficient models that could run on consumer hardware, in contrast to the massive scale of models like GPT-3 (175 billion parameters). Meta trained LLaMA on publicly available data, using a transformer architecture, and released inference code under an open-source license, but kept the weights gated.
Initial Release and Leak
On February 24, 2023, Meta AI announced LLaMA via a blog post and a paper detailing its training and performance. The models came in sizes of 7B, 13B, 33B, and 65B parameters. Meta reported that the 13B model outperformed GPT-3 on many NLP benchmarks, and the 65B model was competitive with state-of-the-art models like PaLM and Chinchilla. Access to weights was granted on a case-by-case basis to academic researchers, government, civil society, and industry research labs.
On March 3, 2023, a torrent containing the weights was uploaded, with the link shared on 4chan and quickly spreading through online AI communities. That same day, a pull request on the official LLaMA repository requested adding the magnet link to documentation. On March 4, another pull request added links to HuggingFace repositories hosting the model. Meta responded by filing takedown requests with HuggingFace on March 6, characterizing the distribution as unauthorized, and HuggingFace complied. On March 20, Meta filed a DMCA takedown against a GitHub repository containing a download script, which was removed the next day.
The leak was met with mixed reactions. Some feared malicious use, such as sophisticated spam or disinformation. Others celebrated the accessibility, noting that smaller models could run on modest hardware, fostering research and innovation. Commentators like Simon Willison compared LLaMA to Stable Diffusion, an open text-to-image model that spurred rapid ecosystem growth. The leak effectively forced Meta to reconsider its release strategy.
Llama 2
On July 18, 2023, Meta, in partnership with Microsoft, announced Llama 2, the next generation. Llama 2 was released in three sizes: 7B, 13B, and 70B parameters. The architecture was largely unchanged from Llama 1, but training data was increased by 40%. Unlike the original, Llama 2 included both foundation models and instruction fine-tuned versions for chat. All models were released with weights and allowed commercial use, subject to an acceptable use policy. This license, however, was not approved by the Open Source Initiative, leading to disputes over whether Llama 2 could be called open source.
Llama 2's release marked a shift toward more open distribution, partly influenced by the leak. It enabled a wave of fine-tuned models and commercial applications. Meta also released Code Llama, a fine-tune of Llama 2 for code generation, with versions on August 24, 2023 (7B, 13B, 34B) and January 29, 2024 (70B).
Llama 3
On April 18, 2024, Meta released Llama 3 with 8B and 70B parameter sizes. These models were pre-trained on approximately 15 trillion tokens of publicly available text, with instruction fine-tuning on over 10 million human-annotated examples. Meta's benchmarks showed Llama 3 70B outperforming Gemini Pro 1.5 and Claude 3 Sonnet on many tasks. Meta announced plans for multilingual and multimodal capabilities, longer context windows, and improved coding and reasoning.
Llama 3 demonstrated that performance continues to scale log-linearly even beyond the Chinchilla-optimal training data amount. For instance, the Chinchilla-optimal dataset for the 8B model was 200 billion tokens, but performance improved up to 15 trillion tokens. During an interview, Mark Zuckerberg noted that the 8B model was nearly as powerful as the largest Llama 2, and the 70B model was still learning at the end of training, but training was stopped to allocate GPU resources elsewhere.
Llama 3.1 was released on July 23, 2024, with further improvements, including a 405B parameter model, and continued the trend of open-weight releases.
Impact on Open-Source AI
The Llama leak catalyzed the open-source AI movement. Before the leak, access to state-of-the-art LLMs was largely limited to a few organizations. The leak allowed independent researchers and hobbyists to run and fine-tune models locally, leading to innovations in model pruning, data augmentation, and efficient inference. It also inspired the creation of open-source ecosystems, such as HuggingFace repositories and tools like llama.cpp, which enabled CPU-based inference.
The leak also pressured Meta to adopt more permissive licenses for subsequent versions, as seen with Llama 2 and Llama 3. This shift influenced other companies, such as Mistral AI, to release open-weight models. However, the debate over what constitutes "open source" in AI continued, with organizations like the Open Source Initiative arguing that restrictions on use and redistribution violate the Open Source Definition.
Reactions and Controversies
Reactions to the leak were polarized. Security experts warned of potential misuse, such as generating phishing emails or disinformation. Some researchers celebrated the democratization of AI, enabling experiments that would otherwise be impossible. The leak also raised questions about the ethics of releasing powerful models, leading to discussions about responsible AI development.
Meta's response to the leak was initially legal, but the company eventually embraced openness. By 2024, Meta positioned Llama as a leading open-weight model family, competing with closed models from OpenAI and Google DeepMind. The leak also highlighted the difficulty of controlling digital distribution, as once weights are public, they cannot be fully retracted.
Legacy
The Llama leak is considered a pivotal event in AI history. It accelerated the development of open-source LLMs, leading to a diverse ecosystem of models and applications. It also influenced policy discussions on AI regulation and the balance between openness and safety. As of 2025, Llama models are widely used in research and industry, and Meta continues to release new versions, including Llama 4 in April 2025. In April 2026, Meta Superintelligence Labs released Muse Spark as a replacement for Llama, marking the next chapter in Meta's AI journey.
The leak demonstrated that open access can drive innovation, but also requires careful consideration of potential harms. It remains a case study in the challenges of governing AI technology in the digital age.