Llama 3 Model Card (GitHub)

Meta Llama 3 is a family of large language models released by Meta AI, available in 8B and 70B parameter sizes as both pre-trained and instruction-tuned variants. The GitHub repository hosting this card has been deprecated following the Llama 3.1 release, with Meta consolidating its infrastructure into a multi-repo Llama Stack. This document represents the original Llama 3 minimal-inference repository README, not a formal model card; the full model card is referenced separately as MODEL_CARD.md.

Both model sizes support text generation and dialogue applications, with a maximum sequence length of 8192 tokens. The 8B model runs on a single GPU (model-parallel value of 1), while the 70B model requires 8-way model parallelism. No benchmark scores or performance metrics are provided in this document.

The card does not discuss evaluation methodology. The only testing-related statement is that "testing conducted to date has not — and could not — cover all scenarios."

The card does not disclose red-team scope, CBRN evaluations, or catastrophic-risk assessments. Meta acknowledges that "Llama 3 is a new technology that carries potential risks with use" and that testing "has not — and could not — cover all scenarios."

Meta directs developers to its Responsible Use Guide for addressing identified risks. The repository notes that additional classifiers can optionally be deployed to filter unsafe inputs and outputs at inference time, with example code provided in the llama-cookbook repository. The PurpleLlama sub-project is identified as the key component of Llama Stack focused on safety risks and inference-time mitigations.

Weights are available via the Meta Llama website and Hugging Face after accepting a license; download links are delivered by email and expire after 24 hours. The models are licensed for both researchers and commercial entities under Meta's custom license and Acceptable Use Policy. Instruction-tuned variants require a specific chat-format prompt structure using special tokens defined in ChatFormat.

Meta states that testing "has not — and could not — cover all scenarios," implying unknown risk surface area. No further technical limitations (hallucination rates, domain gaps, language coverage) are disclosed in this document.

This repository is formally deprecated as of the Llama 3.1 release. Meta has migrated to a consolidated multi-repo Llama Stack spanning llama-models, PurpleLlama, llama-toolchain, llama-agentic-system, and llama-cookbook; users are directed to those repositories going forward.