MENLO PARK, July 12 — It has been nearly two years since Meta released Llama 3.1, the open-source large language model that reshaped the artificial intelligence landscape. For the families and small business owners who now use Meta’s AI assistant across Facebook, Instagram, and WhatsApp, the technology has become a quiet daily companion — helping draft messages, answer questions, and translate between languages. Behind that seamless experience lies a model that, by Meta’s own published evaluations, outperformed GPT-4 on several key benchmarks including MMLU (Massive Multitask Language Understanding) and HumanEval.
Meta released Llama 3.1 on July 23, 2024, in three sizes: 8B, 70B, and 405B parameters. The largest version, with 405 billion parameters, was trained on 16,000 NVIDIA H100 GPUs over 54 days.
Based on GPU rental costs, Meta’s total investment in training Llama 3.1 is estimated at $640 million. The model supports a context length of 128,000 tokens and is multilingual, trained on over 15 trillion tokens of data.
An open alternative to closed models
Meta made Llama 3.1 available for commercial use under a custom license. This open-source release strategy contrasts with closed models from OpenAI and Google. The company also released updated versions of Llama Guard and Code Llama alongside the main model.
As of January 2025, Meta reported that Llama models have been downloaded over 350 million times. The model now powers Meta’s AI assistant integrated into Facebook, Instagram, and WhatsApp.
For users, this means access to a large language model that can handle long documents — up to 128,000 tokens of context — and communicate in multiple languages. The open-source nature of the model has also allowed developers and researchers to build their own applications on top of it, without paying licensing fees to a single corporation.
What the benchmarks mean
Meta’s published evaluations showed Llama 3.1 outperforming GPT-4 on MMLU, a test that measures a model’s ability to answer questions across 57 subjects including law, medicine, and history. The model also scored higher on HumanEval, a benchmark that tests code generation. For the people building tools on top of Llama — from medical diagnosis assistants to educational tutors — those benchmark results signal that an open model can compete with the best proprietary systems.
The $640 million estimated training cost, while enormous, reflects a strategic bet by Meta: invest heavily in a single model, then give it away freely. That bet appears to have paid off in adoption.
With 350 million downloads, Llama has become one of the most widely distributed AI models in the world. The question now is whether Meta can maintain this lead as competitors release their own open models, and as the cost of training ever-larger models continues to climb.
Sources
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Meta releases Llama 3.1, 405B-parameter open-source AI model”, “description”: “MENLO PARK, July 12 — It has been nearly two years since Meta released Llama 3.1, the open-source large language model that reshaped the artificial…”, “datePublished”: “2026-07-12T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “James Roberto”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “citation”: [{“@type”: “CreativeWork”, “url”: “https://en.wikipedia.org/wiki/Llama_%28language_model%29”}]}
