The meta-llama/llama repository is Meta’s official home for Llama model inference code, with nearly 60,000 stars on GitHub — the reference implementation behind the open-weight Llama model family.
When Meta releases Llama models, this repository is where the canonical code lands: model definitions, tokenizers, inference scripts and examples for running the models in pure PyTorch, from small edge-sized checkpoints up to the largest frontier releases. The Llama releases distributed here reshaped the open AI ecosystem — downloadable weights under Meta’s community license enabled the explosion of fine-tunes, quantizations and local-LLM tooling that followed. The repository also hosts generation utilities, chat formatting for instruction-tuned variants and documentation for responsible use. It has trended 40 times in the archive; the code’s last push was in January 2025, reflecting Meta’s cadence of updating it around model releases.
Beyond the code itself, this repository is the distribution point around which the entire open-weights economy organized: fine-tuning frameworks, quantization tools, inference engines and model hubs all track its releases. Each new Llama generation published here resets the baseline for what freely available models can do, from on-device assistants to large-scale research systems.
The code is free from Meta, and Llama model weights are available under Meta’s Llama Community License after registration — free for most research and commercial uses within the license terms.
Researchers and engineers who want the authoritative Llama implementation for study, benchmarking or as the base for their own inference stack.