Definition
Llama is Meta's family of foundation models. The family includes pretrained and instruction-tuned releases, with capabilities and supported modalities that vary by generation. Each Llama release identifies particular model artifacts and terms.
The original name expanded to Large Language Model Meta AI. Later releases use the Llama spelling. Downloadable weights allow operators to run and adapt the released model within its terms; they do not automatically provide all the materials needed to reproduce training.
Origin and attribution
Meta introduced LLaMA on February 24, 2023. Its research team released models to researchers under a noncommercial license. Meta then announced Llama 2 on July 18, 2023, including pretrained and conversationally tuned models with research and commercial access under stated terms.
Meta's research teams developed the family; the original announcement records their release and its name.
Disagreement over openness
Meta described Llama 2 as open source in its launch announcement. On July 20, 2023, Open Source Initiative executive director Stefano Maffulli disputed that characterization. OSI pointed to restrictions on some commercial users and permitted purposes, which conflict with the Open Source Definition's nondiscrimination requirements.
The disagreement concerns the rights and artifacts required for the open-source label. OSI's later Open Source AI Definition 1.0 also requires data information, code, and parameters in the preferred form for modification. Weight availability alone is therefore insufficient for that standard. Open-weight describes the released parameters without resolving that broader claim.
Scope and operational significance
Licenses need a release-specific reading. The Llama 4 Community License, for example, contains attribution requirements, an acceptable-use policy, and additional commercial terms for organizations above its stated scale threshold. Those terms cannot be replaced with assumptions drawn from the original research release or from another vendor's model.
Preserve lineage when using fine-tunes or distilled models based on Llama. A different publisher name does not remove the need to examine the source model's terms. Likewise, quantization and instruction tuning can change behavior while retaining the underlying lineage.
Evaluate the actual artifacts and serving setup on the intended workload. Local access gives an operator control over deployment, while also making that operator responsible for serving, updates, provenance checks, and runtime validation.
Distinguish it from nearby terms
- Open-weight describes access to trained parameters. Open-source AI adds rights and artifacts that are the subject of the named dispute.
- A foundation model is a reusable model class; Llama identifies a vendor family within it.
- Fine-tuning modifies a pretrained model for a task or behavior. A Llama-derived fine-tune retains a lineage the deployment must track.
Check your understanding
A model card calls a Llama-derived model open source because its weights can be downloaded. Which license terms, training materials, and lineage evidence would you need to assess that claim?