Llama

Llama is Meta's family of language models for developers, researchers and businesses building AI applications on their own hardware or servers. Downloadable weights let you choose a model for your application and available hardware.

• Text completion and chat: Base models support text completion, while instruction-tuned models support conversational applications.
• Vision models: The family includes models that work with images as well as smaller models for text applications.
• Lower memory requirements: FP8 and Int4 quantization reduce the memory needed for inference, with some loss of accuracy.
• Runtime support: Hugging Face distributes weights for Transformers and in native Llama 4 format, while Llama Stack also supports inference through other providers.

Access to the weights requires license acceptance and approval. The weights use custom licensing for research and commercial use. The Python toolset includes download utilities and inference examples.

Hardware needs depend on the model and precision. Full-precision (bf16) Llama 4 inference requires at least four GPUs. For Llama-4-Scout-17B-16E-Instruct specifically, FP8 inference requires two GPUs with 80 GB of memory each; Int4 requires one GPU with 80 GB of memory.

The official model repository distinguishes Llama 2, 3, 3.1, 3.2, 3.3 and 4 releases and links a license for each generation. This listing includes Llama 4 Scout and Maverick; those weights use the Llama 4 Community License and acceptable-use policy, not a permissive open-source code license.