What it does

This repository collects utilities for working with the Llama family of large language models. Its README tabulates releases from Llama 2 (July 2023) through Llama 4 (April 2025), listing sizes, context lengths, tokenizers, use policies, licenses and model cards for each generation. To get weights, you request access on the Llama website, accept the license and receive a signed download URL by email, which stops working after a day or a set number of downloads. A llama-model command-line tool installed with pip can list, describe, download, verify and remove models and show a model's prompt format. Weights can also be fetched from Hugging Face after accepting the license. Example scripts run chat and text completion with torchrun once the torch extras are installed. The README says the Llama 4 series needs at least four GPUs at full bf16 precision, and that FP8 or Int4 mixed-precision modes reduce this: Llama 4 Scout Instruct runs on two 80GB GPUs with FP8, or on a single 80GB GPU with Int4. For wider inference options it points to the Llama Stack toolset. The README also covers Hugging Face downloads in transformers and native formats, a Responsible Use Guide, and routes for reporting model issues and risky generated content.

Capabilities

Release table covering Llama 2, 3, 3.1, 3.2, 3.2-Vision, 3.3 and 4 with sizes and context lengths
llama-model CLI to list, describe, download, verify and remove models
Example chat and completion scripts launched with torchrun
FP8 and Int4 mixed-precision modes to reduce GPU memory needs
Weights also available through Hugging Face in transformers and native formats
Installable as a Python package with pip install llama-models

When to use it

You want to download Llama weights and run them yourself on GPU hardware
You need prompt formats and model details for a specific Llama release
You are testing lower-memory FP8 or Int4 inference for a Llama 4 model