Llama (Large Language Model Meta AI) is a family of large language models developed by Meta AI, first released on February 24, 2023. It was positioned as a research-oriented alternative to closed commercial models, made available with published weights under a license permitting research and, in later versions, most commercial use.
Initial release and leak
The first Llama models, ranging from 7 billion to 65 billion parameters, were initially released only to approved researchers under application. Within days, the model weights leaked onto the file-sharing site 4chan and spread rapidly across the internet, an event widely credited with jump-starting a wave of grassroots experimentation with running and fine-tuning capable language models outside major labs. The leak, combined with techniques like Quantization that made it feasible to run large models on consumer hardware, enabled a surge of community projects and fine-tuned derivatives built on the Llama weights, hosted extensively on platforms like Hugging Face.
Subsequent releases
Meta followed with Llama 2 in July 2023, released in partnership with Microsoft under a more permissive license that explicitly allowed commercial use for most organizations, and Llama 3 in April 2024, which improved benchmark performance substantially and was later extended with larger and multimodal variants, including versions accepting image input. Llama 4, released in 2025, adopted a mixture-of-experts architecture and native multimodality. Across versions, Yann LeCun, Meta's chief AI scientist at the time, was a prominent public advocate for the strategic value of releasing model weights openly, arguing it accelerated safety research, prevented dangerous concentration of AI capability, and built developer ecosystems around Meta's technology stack.
Licensing and the open-weights debate
Llama's license terms evolved across versions and were a subject of ongoing debate about whether the models qualified as genuinely "open source" in the traditional software sense, since the license included use restrictions (such as limits for very large companies) not present in standard open-source licenses. This tension placed Llama at the center of broader industry discussion about open-weights models and what openness should mean for AI, alongside other open releases from Mistral AI, DeepSeek, and Alibaba's Qwen.
Ecosystem and impact
Llama's availability significantly shaped the broader open-model ecosystem: it became a common base for LoRA and other efficient fine-tuning methods, a standard baseline in academic research on model behavior, and the foundation for many downstream commercial and open products. Mark Zuckerberg publicly framed Meta's open-weights strategy as a long-term bet distinct from the closed approaches of OpenAI and other frontier labs, arguing that broad access to capable models served both Meta's business interests and the wider AI ecosystem, though the strategy drew criticism from some AI safety researchers concerned about the difficulty of controlling misuse once weights are freely distributed. In 2025, Meta reorganized parts of its AI research effort under a new "superintelligence" lab, recruiting external talent including Alexandr Wang, amid reports of an increasingly aggressive, less openly published research posture relative to the original open Llama releases.