Llama (stylized as LLaMA, with "Large Language Model Meta AI" as a backronym) is a family of large language models (LLMs) developed by Meta AI, first released in February 2023. The models range in size from 1 billion to 2 trillion parameters. Initially, the first version was a foundation model available only to researchers under a non-commercial license, but starting with Llama 2, Meta released instruction fine-tuned versions alongside foundation models, with broader licensing. The unauthorized leak of the first model's weights in March 2023 via BitTorrent significantly accelerated the proliferation of open-source LLM derivatives and tools.
Llama models are based on the Transformer (architecture) architecture, similar to other contemporary LLMs. They have been released in multiple versions, each introducing improvements in scale, training data, and capabilities. As of 2025, the latest version is Llama 4, released in April 2025, and Meta has also integrated Llama into its Meta AI assistant.
Background
Following the release of earlier large language models such as GPT-3, a major research focus was up-scaling models, which in some cases led to significant gains in emergent capabilities. The launch of ChatGPT in late 2022 and its unexpected success dramatically increased public and industry attention on LLMs. In response, Meta's Chief AI scientist Yann LeCun commented that large language models are best suited for aiding with writing, reflecting a cautious stance compared to other tech companies' aggressive deployment.
Initial Release
The first version of Llama (sometimes referred to as Llama 1) was announced on February 24, 2023, via a blog post and a paper detailing its training, architecture, and performance. The inference code was released under the open-source GPLv3 license, but access to the model weights was gated behind an application process, with access granted on a case-by-case basis to academic researchers, government, civil society, and industry research laboratories. The model was trained exclusively on publicly available data, with various model sizes (7B, 13B, 33B, and 65B parameters) to accommodate different hardware capabilities. It was a foundation model only, though the paper included examples of instruction fine-tuning.
Meta reported that the 13B parameter model outperformed the much larger GPT-3 (175B parameters) on most NLP benchmarks, and the largest 65B model was competitive with state-of-the-art models like PaLM and Chinchilla.
Leak and Aftermath
On March 3, 2023, a torrent containing Llama's weights was uploaded, with a link shared on the 4chan imageboard and subsequently spread 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 filed takedown requests on March 6, characterizing the distribution as unauthorized, and HuggingFace complied. On March 20, Meta filed a DMCA takedown against a repository containing a script to download Llama from a mirror, and GitHub complied the next day.
Reactions to the leak were mixed. Some speculated the model could be used for malicious purposes, such as more sophisticated spam. Others celebrated the accessibility, noting that smaller versions could run relatively cheaply, potentially fostering further research. Commentators like Simon Willison compared Llama to Stable Diffusion, an openly distributed text-to-image model that spurred rapid tool development, suggesting a similar effect for LLMs.
Llama 2
On July 18, 2023, in partnership with Microsoft, Meta announced Llama 2, the next generation, with model sizes of 7B, 13B, and 70B parameters. The architecture remained largely unchanged from Llama 1, but training used 40% more data. Llama 2 included both foundation and chat fine-tuned models, and all weights were released for many commercial uses. However, because the license includes an acceptable use policy, it is not considered open source by the Open Source Initiative, and Meta's use of the term has been disputed.
Code Llama, a fine-tune of Llama 2 on code-specific datasets, was released in versions: 7B, 13B, and 34B on August 24, 2023, and a 70B version on January 29, 2024. The foundation models were trained on an additional 500B tokens of code data, then 20B tokens of long-context data, with further instruction tuning on 5B tokens. A Python-specific model was trained on 100B tokens of Python code.
Llama 3
On April 18, 2024, Meta released Llama 3 in two sizes: 8B and 70B parameters. The models were pre-trained on approximately 15 trillion tokens from publicly available sources, with instruct models fine-tuned on public instruction datasets and over 10 million human-annotated examples. Meta's testing showed that Llama 3 70B outperformed Gemini Pro 1.5 and Claude 3 Sonnet on most benchmarks. Meta announced plans to make Llama 3 multilingual, multimodal, better at coding and reasoning, and to increase its context window.
Regarding scaling laws, Llama 3 models empirically showed that performance continues to scale log-linearly even beyond the Chinchilla-optimal amount of training data. For example, the Chinchilla-optimal dataset for Llama 3 8B is 200 billion tokens, but performance improved up to 15 trillion tokens (75 times more). In an interview, Mark Zuckerberg noted that the 8B version was nearly as powerful as the largest Llama 2, and that the 70B model was still learning at the end of training, which was stopped to allocate GPU resources elsewhere.
Llama 3.1 was released on July 23, 2024, introducing a 405B parameter model, along with updated 8B and 70B versions, with a longer context window of 128K tokens and improved multilingual support.
Llama 4 and Beyond
Llama 4 was released in April 2025, featuring a mixture-of-experts architecture with sizes up to 2 trillion parameters. It introduced multimodal capabilities and improved reasoning. In April 2026, Meta Superintelligence Labs released Muse Spark as a replacement for Llama, signaling a shift in Meta's AI strategy.
Impact and Ecosystem
The leak of Llama 1 weights catalyzed a vibrant open-source ecosystem, with numerous fine-tunes and derivatives emerging. It enabled researchers and hobbyists to experiment with LLMs without API costs, leading to innovations in quantization, efficient inference, and local deployment. The availability of smaller models like 7B and 13B made it feasible to run LLMs on consumer hardware, democratizing access. This contrasted with the closed approaches of other major labs like OpenAI and Anthropic, and influenced subsequent releases from other organizations. The controversy over licensing also highlighted tensions between open-source ideals and corporate control, a debate that continues in the AI community.