The llama-models repository serves as Meta's official utilities and distribution system for Llama large language models. Written primarily in Python, it provides developers, researchers, and businesses with tools to download, configure, and run various versions of the Llama model family. The repository hosts documentation and utilities for multiple Llama generations, including Llama 2, Llama 3, Llama 3.1, Llama 3.2, Llama 3.2-Vision, Llama 3.3, and Llama 4, with model sizes ranging from 1B to 405B parameters and context lengths extending up to 128K tokens or 10M tokens for the newest Llama 4 Scout variant.
The primary feature of this repository is the llama-models CLI tool, installable via pip, which streamlines the model download and management process. Users can list available models using the llama-model list command, with options to view all historical versions, and download specific model checkpoints using signed URLs obtained from the Meta Llama website. The CLI handles authentication and manages the time-limited download links that expire after 24 hours. The repository also provides example scripts for running inference with different model variants, including separate scripts for instruct and base models, with specific guidance that Llama 4 series models require at least 4 GPUs for full precision inference.
A significant capability documented in the repository is quantization support for reducing memory requirements. Users can run inference with FP8 mixed precision or Int4 mixed precision modes, substantially lowering the GPU memory footprint needed. For instance, the Llama 4 Scout 17B model can run on 2 GPUs with 80GB memory using FP8 quantization or a single 80GB GPU with Int4 quantization. The repository also integrates with Hugging Face, offering alternative download paths through the meta-llama organization on that platform, with support for both transformers and native llama4 formats.
Community engagement around this repository shows consistent activity. GitGenius tracking reveals 223 issues and pull requests with a median response latency of 24.1 hours, indicating active maintenance. The most active contributors tracked include ashwinb with 101 events, varunfb with 41 events, and raghotham with 29 events. The llama api label appears most frequently in issue tracking with 3 occurrences. The repository has grown modestly, with fork count increasing from 1397 to 1398 between tracking periods. The codebase connects to major machine learning ecosystems, with overlapping contributors shared with pytorch/pytorch, huggingface/transformers, and tensorflow/tensorflow repositories.
The repository emphasizes responsible AI development, including detailed model cards, acceptable use policies, and licensing information for each model version. It directs users to a Responsible Use Guide and provides multiple channels for reporting issues, including bug reports through GitHub issues, content safety concerns through Facebook's developer feedback system, and security vulnerabilities through Facebook's whitehat program. The documentation explicitly notes that model weights are licensed for both researchers and commercial entities, supporting Meta's stated mission of fostering open access to large language models while maintaining trust and safety standards.
