Llama 3 Release, License, and Lifecycle Record

Llama 3 Release, License, and Lifecycle Record

Meta introduced Llama 3 on April 18, 2024. The first release contained pretrained and instruction-tuned 8B and 70B text models with an 8,192-token context.

Meta’s launch record reports more than 15 trillion training tokens from publicly available sources, four times as much code as Llama 2, grouped-query attention, a 128K-token tokenizer vocabulary, vendor benchmark results, and a separate human-evaluation set. These are Meta disclosures and evaluation claims.

The official model card identifies the April 18 release, English intended use, static offline training data, supported variants, evaluation results, limitations, and the custom license. The exact checkpoint and intended-use statement matter when describing a result.

The community license grants use, reproduction, distribution, modification, and derivative-work rights with conditions. It requires notices and branding in defined redistribution cases, incorporates an acceptable-use policy, restricts using the materials or outputs to improve another large language model, and adds terms for products with more than 700 million monthly active users.

Meta described models above 400B parameters as still training in April. Llama 3.1 later introduced 405B and updated 8B and 70B models on July 23, 2024 with a longer context and other changes. The 405B model must not be projected backward into E013.

The official Llama models repository now distinguishes Llama 3, 3.1, 3.2, 3.3, and 4 by launch date, size, context, policy, license, and model card. Llama 3 remains a historical release, not a synonym for every later Llama model or every Meta AI experience.