LLaMA 2 (stylized as LLaMa 2) is a family of large language models released by Meta AI on July 18, 2023, in partnership with Microsoft. It succeeded the original LLaMA model and was notable for making model weights available for commercial use, a departure from the earlier version's restricted academic-only access. The release included both foundation models and instruction fine-tuned chat variants, with sizes of 7, 13, and 70 billion parameters.
The model architecture remained largely unchanged from LLaMA 1, but training data was increased by 40%. LLaMA 2 was trained on publicly available data, and the chat versions were fine-tuned using techniques such as RLHF (reinforcement learning from human feedback). The models were distributed under a custom license that permitted commercial use but imposed an acceptable use policy, leading to disputes over whether they qualified as open-source software.
Background
The release of ChatGPT in late 2022 and its rapid popularity intensified interest in generative AI and large language models. Meta's chief AI scientist, Yann LeCun, had previously stated that such models were best suited for aiding writing, positioning Meta's approach within the broader AI landscape.
LLaMA 1, announced on February 24, 2023, was a foundation model family ranging from 1B to 65B parameters. Its weights were initially available only to researchers on a case-by-case basis under a non-commercial license. In March 2023, unauthorized copies of LLaMA 1 were leaked via BitTorrent and HuggingFace, prompting Meta to file takedown requests. The leak was compared to the open distribution of Stable Diffusion, which had spurred rapid community innovation.
Release and Features
LLaMA 2 was announced on July 18, 2023, with three parameter sizes: 7B, 13B, and 70B. The foundation models were trained on 2 trillion tokens of publicly available data, a 40% increase over LLaMA 1. The chat models were fine-tuned on over 1 million human annotations and used RLHF to align responses with user intent.
The license for LLaMA 2 allowed commercial use, but it included restrictions such as prohibiting use to improve other large language models. This led the Open Source Initiative to argue that LLaMA 2 was not truly open source, despite Meta's use of the term.
Commercial and Ecosystem Impact
LLaMA 2's commercial availability made it a popular choice for startups and enterprises seeking an alternative to proprietary models from OpenAI and Anthropic. It was integrated into Amazon Web Services (AWS), Azure, and Google Cloud platforms, enabling deployment via managed services. The 70B model was particularly noted for its performance on benchmarks, often rivaling larger proprietary models.
In August 2023, Meta released Code Llama, a fine-tuned version of LLaMA 2 specialized for code generation, with sizes 7B, 13B, 34B, and later 70B. This expanded LLaMA 2's applicability to software development tasks.
Successors and Legacy
LLaMA 2 was succeeded by LLaMA 3 on April 18, 2024, which introduced 8B and 70B models trained on 15 trillion tokens. LLaMA 3 demonstrated that performance continued to improve even beyond the Chinchilla-optimal training data amount, a finding that influenced scaling practices. LLaMA 3.1 followed on July 23, 2024, with a 405B parameter model, and LLaMA 4 was released in April 2025.
In April 2026, Meta Superintelligence Labs introduced Muse Spark as a replacement for LLaMA, marking a shift in Meta's AI model strategy. LLaMA 2 remains a significant milestone in the history of open-weight large language models, having set a precedent for commercial accessibility and shaping the ecosystem of generative AI.