Llama 2
Meta's open family of pretrained and chat-tuned language models, 7B to 70B
Llama 2 is a family of pretrained and dialogue-tuned large language models released by Meta in July 2023, with sizes from 7 to 70 billion parameters. The chat versions, called Llama 2-Chat, were made free for research and commercial use, with weights and starting code published alongside a safety-focused paper.
- Pretrained and chat-tuned models from 7 billion to 70 billion parameters
- Weights and starting code released free for research and commercial use
- Published on arXiv on 18 July 2023, with Thomas Scialom submitting
- Microsoft named preferred partner; offered in the Azure AI model catalog
- Red-teamed internally and by commissioned third parties
What it is
Llama 2 is a set of large language models built and released by Meta. It covers pretrained models and dialogue-tuned variants named Llama 2-Chat, at scales from 7 billion to 70 billion parameters. The chat models are aimed at conversational use. Meta released the weights and starting code for both the pretrained and the conversational versions, free for research and commercial use. The company positions the family as an open alternative to closed models, and says its fine-tuned versions beat open-source chat models on most benchmarks it tested and, on its human ratings for helpfulness and safety, could stand in for closed-source systems.
How it came about
Llama 2 followed Llama 1, which Meta says drew more than 100,000 access requests from researchers. The work was published as an arXiv paper on 18 July 2023, revised the next day, with Hugo Touvron and Louis Martin among the lead authors and Thomas Scialom submitting it. Meta announced availability the same day, with Microsoft named as preferred partner. Satya Nadella presented the partnership at Microsoft Inspire. Meta also launched an academic research community programme and a Llama Impact Challenge around the release.
Inside the model
The paper describes Llama 2 as pretrained and then fine-tuned for dialogue, with the tuned models called Llama 2-Chat. Meta says the fine-tuning and safety work is set out in detail so others can build on it, including a transparency schematic inside the paper that lists known problems and the mitigations tried. Safety fine-tuning used adversarial prompts generated by the team, plus external adversarial testing commissioned from third parties. Meta describes this process as iterative, with updated tuned models planned as benchmarking continues. Specifics such as layer counts, attention design, training data and compute are not given in the available material.
What it can do
Meta reports that its fine-tuned models outperform open-source chat models on most benchmarks it tested. Based on its own human evaluations of helpfulness and safety, the company says Llama 2-Chat may serve as a substitute for closed-source models. The company also acknowledges shortcomings in the models and sets some of these out in the transparency schematic within the research paper. No individual benchmark scores appear in the available material.
How to use it
Llama 2 is free for research and commercial use, with model weights and starting code released for the pretrained and conversational versions. It is offered in the Azure AI model catalog, optimised to run locally on Windows, and available through Amazon Web Services, Hugging Face and other providers. An Acceptable Use Policy bans certain use cases, and a Responsible Use Guide sets out recommended practices for developers.
Safety and testing
Meta says the fine-tuned models were red-teamed through internal work and external efforts, with the team producing adversarial prompts to aid fine-tuning and third parties commissioned for adversarial testing. Safety fine-tuning is described as ongoing, with further tuned models planned. Supporting resources include a transparency schematic in the research paper covering known issues and mitigations, a Responsible Use Guide for developers, and an Acceptable Use Policy restricting certain uses.
Reception and debate
Meta says demand for the earlier Llama 1 was strong, with more than 100,000 access requests from researchers, and describes broad support for the open approach behind Llama 2 from companies, cloud providers, research institutions and people in tech, academia and policy. The available material does not record outside reviews, criticism, lawsuits or bans.
Why it matters
Meta frames Llama 2 as part of a longer push at the company for open research and collaboration, and says open access lets a wide community stress-test models and find problems quickly. The release was backed by cloud and platform partners, including Microsoft as preferred partner, and by new programmes meant to gather outside feedback: an Open Innovation AI Research Community for academics and a Llama Impact Challenge inviting public, non-profit and for-profit groups to apply the models to environmental, education and other problems.
Questions people ask about Llama 2
What is Llama 2?
A family of large language models from Meta, released in July 2023, running from 7 to 70 billion parameters. It includes pretrained models and dialogue-tuned versions called Llama 2-Chat, with weights and starting code published for both.
Can I use it commercially?
Yes. Meta made Llama 2 free for research and commercial use, and released the model weights and starting code. An Acceptable Use Policy prohibits certain use cases, and a Responsible Use Guide sets out recommended practices.
How does it compare to other models?
Meta says its fine-tuned models beat open-source chat models on most benchmarks it tested, and that human evaluations for helpfulness and safety suggest they may substitute for closed-source models. Individual benchmark scores are not given in the available material.
What safety work was done?
The fine-tuned models were red-teamed internally and by commissioned third parties, with adversarial prompts used during fine-tuning. Meta describes the process as iterative and published a transparency schematic covering known issues and mitigations.
Where can I get it?
Through the Azure AI model catalog, Amazon Web Services, Hugging Face and other providers, and optimised to run locally on Windows. Weights and starting code are included for the pretrained and conversational versions.
Good reading and sources
There is no English Wikipedia article on Llama 2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2307.09288, about.fb.com/news/2023/07/llama-2/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.