Meta Llama 3: Release, Models, License, and Lifecycle

Meta Llama 3 Release and License Record

Meta Llama 3 is the model family Meta introduced on April 18, 2024. The initial release included pretrained and instruction-tuned text models with 8 billion and 70 billion parameters.

The larger model discussed in E013 was still training. It was not part of the April release.

timeline
    title Llama 3 release boundary
    2024-04-18 : Llama 3 8B and 70B released
    2024-04-23 : Venture Step E013 published
    2024-07-23 : Llama 3.1 8B, 70B, and 405B released
    2024-09-25 : Llama 3.2 releases followed
    2024-12-04 : Llama 3.3 70B followed

What Meta released in April 2024

Meta’s Llama 3 launch record described 8B and 70B pretrained and instruction-tuned models available for broad use through Meta and distribution partners.

The company reported an 8,192-token context, a tokenizer vocabulary of 128,000 tokens, grouped-query attention, and more than 15 trillion pretraining tokens collected from publicly available sources. Meta also said the training set was seven times larger than Llama 2’s and contained four times as much code.

Those are vendor disclosures. They establish what Meta reported about the release. They do not independently establish data completeness, task performance, safety, or production fitness.

The official model card identifies the model variants, English intended use, static offline training cutoff information, evaluation results, limitations, and the custom license. It is the better source for exact model identity than a general product announcement.

What remained in training

Meta said its largest Llama 3 models exceeded 400 billion parameters and were still training.

The surviving E013 outline refers to a forthcoming model as 408B. That was not an available model on the episode date. Meta later released Llama 3.1 405B on July 23, 2024 alongside updated 8B and 70B models.

The later release also expanded context to 128K, added multilingual support and tool-use claims, and changed parts of the license. Those details belong to Llama 3.1. They should not be projected backward into the April Llama 3 release.

The license

The Meta Llama 3 Community License granted a non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, modify, and create derivative works from the Llama materials.

The license also imposed conditions. Defined redistribution required the agreement, notices, and “Built with Meta Llama 3” branding. Use had to follow the acceptable-use policy. The agreement restricted using the materials or outputs to improve another large language model. Products or services above 700 million monthly active users had additional commercial terms.

That combination supports broad use. It is not the same as unrestricted access.

Meta described Llama 3 as open source. The Open Source Initiative later concluded that Llama 3.x did not meet its standards. Venture Step therefore uses open-weight or openly available under Meta’s license unless directly attributing Meta’s wording.

This is descriptive language, not a legal conclusion about a particular use.

Benchmark boundaries

Meta published automatic benchmark results and a separate human-evaluation comparison. The launch page tells readers to inspect the evaluation settings.

A score belongs to the exact model variant, prompt and shot setup, scoring method, implementation, and date. It does not prove that a model will be best for a user’s task, cheaper to operate, safer in an application, or more reliable after tools and retrieval are added.

The [[How to Read an MMLU Score|MMLU explainer]] shows how a broad average can hide subject variation and method differences.

Current lifecycle

The official Llama models repository now separates Llama 3, 3.1, 3.2, 3.3, and 4 by release date, size, context, license, policy, and model card.

That record makes Llama 3 a historical model family. It should not be used as a current umbrella for every Llama release or every Meta AI experience.

The historical record still matters because applications, research, archived evaluations, and compliance decisions can depend on the exact artifact originally adopted. Replacing the family name with a newer label does not revise the license, model card, or operating assumptions attached to an older deployment.

An adopter should identify the exact checkpoint, source, license, intended use, hardware path, safety tools, evaluation setup, maintenance owner, and retirement plan. A family name is not an operating specification.

E026 and E027 continue the Llama 3.1 story. E029 examines training and benchmark claims. E031 focuses on safety testing and inference. Together they should form a dated model-history cluster rather than one timeless profile.

The maintained [[Meta Llama 3 Model Profile]] owns the canonical historical identity. This E013 record owns the release, license, and later-model boundary. AI assisted with research, drafting, and validation. Publication remains unauthorized pending model, technical, license, source, accessibility, and founder review.