Two sizes, a much bigger training run, and a third still cooking
Meta released Llama 3 on April 18, 2024, in pretrained and instruction-tuned versions at 8B and 70B parameters. The models were trained on over 15 trillion tokens, about seven times the Llama 2 training set and with four times as much code, using a 128K-token vocabulary the company said cut tokenized length by up to 15 percent compared to Llama 2. Context length stayed at 8,192 tokens, unchanged from the prior generation. [1]
Meta said its "largest models are over 400B parameters" and were "still training" as of the release, with plans for multimodality, multilingual conversation and a longer context window once that model shipped. [1]
What '8B and 70B' scored against paid models
Meta's model card reports the instruction-tuned 70B at 82.0 on MMLU (5-shot), 81.7 on HumanEval (0-shot), 93.0 on GSM8K (8-shot, chain of thought), 50.4 on MATH (4-shot, chain of thought) and 39.5 on GPQA (0-shot). The smaller instruction-tuned 8B scored 68.4, 62.2, 79.6, 30.0 and 34.2 on the same five benchmarks. Meta's own blog post called the pair "the best models existing today at the 8B and 70B parameter scale," and its release charts placed 70B against Claude 3 Sonnet, Mistral Medium and Gemini Pro 1.5 rather than only against smaller open models. [2][1]
Atlas interpretation: The summary above says these weights "held their own against paid frontier models on most public evaluations." That is a claim about relative standing at release, not a claim that either model matched a frontier system's ceiling: a 70B open-weight model scoring near mid-tier closed models on public benchmarks is a genuinely different competitive position than Llama 2 held eight months earlier, when its comparisons were pitched mostly against other open models rather than paid ones. [2][1]
The same weights, wired into a consumer assistant
Meta paired the model release with an upgrade to Meta AI, describing it as "built with Llama 3" and now running across Facebook, Instagram, WhatsApp, Messenger and a standalone website. That distinguishes Llama 3 from a research artifact released for others to build on: the same weights were, on the same day, put in front of a consumer user base that dwarfs the audience any single Llama 2 deployment reached. [1]
Meta also introduced Llama Guard 2, Code Shield and CyberSec Eval 2 alongside the release, framed as tools for developers to filter unsafe inputs and outputs and to check generated code, and said the models had been red-teamed through internal and external testing before release. [1]
The same license terms, the same dispute unresolved
Llama 3 shipped under the Meta Llama 3 Community License Agreement, which repeats the structure the Open Source Initiative had already objected to at the Llama 2 release. A licensee whose products had more than 700 million monthly active users in the preceding month must first request a separate license from Meta, and the acceptable use policy still bars using the Llama Materials or their outputs "to improve any other large language model," Llama 3 and its derivatives excepted. Redistribution requires prominently displaying "Built with Meta Llama 3" and retaining Meta's license and copyright notice. [3]
Atlas interpretation: That the terms are functionally unchanged matters for the event's "open-source" category label the same way it did for the prior release: a size-gated permission requirement and a restriction on which models a licensee may improve are still not features of a permissive open-source license, whatever the weights' availability and Meta's own description of them. Nothing about the wider benchmark gap changes that classification question; it only raises the stakes of who is asking to use the weights. [3]
Sources
- Introducing Meta Llama 3: The most capable openly available LLM to date
Meta AI · Apr 18, 2024
- meta-llama/Meta-Llama-3-70B-Instruct
Hugging Face · Sep 8, 2026
- llama3/LICENSE
Meta · Apr 18, 2024