Llama 3

Meta's April 2024 Llama 3 release, in 8B and 70B sizes with base and instruct variants, raised expectations for what open-weight models could do in chat and coding. It brought a 128K-entry tokenizer vocabulary, grouped-query attention in both sizes, and pretraining on more than 15 trillion tokens, though its context window was only 8K tokens and Meta replaced it within three months. The meta-llama/llama3 repository, a minimal PyTorch inference example plus a download script, was deprecated with the Llama 3.1 launch in July 2024 and is now archived on GitHub, with its last commit in January 2025. The weights are still available from Hugging Face after accepting the Meta Llama 3 Community License, which allows commercial use but requires attribution and a separate license for services above 700 million monthly users, and they run in llama.cpp, Ollama, vLLM and Transformers. For new work, start from Llama 3.1 or 3.3 with 128K context through the llama-models repository, or from Llama 4 Scout and Maverick, and compare them against other current open-weight families before committing.