What is Llama? Meta AI’s family of large language models explained

Llama, a family of sort-of open-source large language models (LLMs), was introduced in February 2023, and has been updated periodically ever since. Llama shook up the world of LLMs, which had previously been dominated by ever-larger closed-source models from OpenAI and Google. With Llama, Meta AI touted the fact that smaller generic models trained on more tokens are easier and cheaper to use for retraining and fine-tuning specialized models than huge models.

Why do I say “sort-of open-source?” It’s because the Meta Llama license has restrictions on commercial use (aimed at large, for-profit organizations, such as AWS, Google Cloud, and Microsoft Azure) and acceptable use (forbidding it, for example, from being used to develop weapons or drugs). They’re reasonable restrictions, but the Open Source Initiative maintains that they violate the official definition of open source.

I discussed the Llama 2 models in September 2023. I also explored the use of Llama Chat and Code Llama for code generation.

Meta Llama lawsuits

Two class action lawsuits were filed by authors against Meta in 2023 claiming misuse of their copyrighted books as training material for Llama. The courts haven’t been terribly sympathetic to the authors. The first case, Kadrey et al. v. Meta Platforms, was filed in July 2023. The second, Chabon v. Meta Platforms, was filed in September 2023. The two cases were combined, and the combined case was dismissed on all but one count, direct copyright violation, in December 2023. It still dragged on with amended complaints for another year; in September 2024 the judge ordered “Based on previous filings, the Court has been under the impression that there’s no real dispute about whether Meta fed copyrighted works to its AI programs without authorization, and that the only real legal question to be presented at summary judgment is whether doing so constituted ‘fair use’ within the meaning of copyright law.” Summary judgement was scheduled for March 2025, although there has been plenty of activity from both sides through February 2025.

Meta Llama models (since Llama 2)

Below I’ll discuss the progress Meta AI has made with the Llama family of models since the fall of 2023. Note that they’re no longer just language models. Some models are multi-modal (text and vision inputs, text output), and some can interpret code and call tools. In addition, some Llama models are safety components that identify and mitigate attacks, designed to be used as part of a Llama Stack. The following model write-ups are condensed from Meta AI’s model cards.

Llama Guard 1

Llama Guard is a 7B parameter Llama 2-based input-output safeguard model. It can be used for classifying content in both LLM inputs (prompt classification) and in LLM responses (response classification).

The six-level taxonomy of harms used by Llama Guard is Violence & Hate; Sexual Content; Guns & Illegal Weapons; Regulated or Controlled Substances; Suicide & Self Harm; and Criminal Planning.

Model release date: December 13, 2023

Code Llama 70B

With the introduction of Code Llama 70B, Code Llama comes in three variants:

Code Llama: Base models designed for general code synthesis and understanding.

Code Llama – Python: Models designed specifically for Python.

Code Llama – Instruct: Models designed for following instructions and safer deployment.

All variants are available in sizes of 7B, 13B, 34B, and 70B parameters. Code Llama and its variants are intended for commercial and research use in English and relevant programming languages. Code Llama is indeed good at coding; see my review.

Model release date: January 29, 2024

Llama Guard 2

Llama Guard 2 is an 8B parameter Llama 3-based LLM safeguard model. The model is trained to predict safety labels on 11 categories, based on the MLCommons taxonomy of hazards.

Hazard categories

S1: Violent CrimesS2: Non-Violent CrimesS3: Sex-Related CrimesS4: Child Sexual ExploitationS5: Specialized AdviceS6: PrivacyS7: Intellectual PropertyS8: Indiscriminate WeaponsS9: HateS10: Suicide & Self-HarmS11: Sexual Content

Model release date: April 18, 2024

Meta Llama 3

Meta Llama 3 comes in two sizes, 8B and 70B parameters, in pre-trained and instruction-tuned variants. Instruction-tuned models are optimized for dialogue.

Model release date: April 18, 2024

Prompt Guard

Prompt Guard is a classifier model that is capable of detecting both explicitly malicious prompts (jailbreaks) and prompts that contain injected inputs (prompt injections). Meta AI suggests fine-tuning the model to application-specific data to achieve optimal results.

Prompt Guard is a BERT model that outputs only labels. Unlike Llama Guard, Prompt Guard doesn’t need a specific prompt structure or configuration. The input is a string that the model labels as safe or unsafe (at two different levels).

Model release date: July 23, 2024

Llama Guard 3

Llama Guard 3 comes in three flavors: Llama Guard 3 1B, Llama Guard 3 8B and Llama Guard 3 11B-Vision. The first two models are text only. The third supports the same vision understanding capabilities as the base Llama 3.2 11B-Vision model. All the models are multi-lingual (for text-only prompts) and follow the categories defined by the MLCommons consortium.

You can use Llama Guard 3 8B for category S14 Code Interpreter Abuse. Note the Llama Guard 3 1B model was not optimized for this category.

Hazard categories

S1: Violent CrimesS2: Non-Violent CrimesS3: Sex-Related CrimesS4: Child Sexual ExploitationS5: DefamationS6: Specialized AdviceS7: PrivacyS8: Intellectual PropertyS9: Indiscriminate WeaponsS10: HateS11: Suicide & Self-HarmS12: Sexual ContentS13: ElectionsS14: Code Interpreter Abuse

Model release date: July 23, 2024

Meta Llama 3.1

The Meta Llama 3.1 collection of multi-lingual large language models includes pre-trained and instruction-tuned generative models in 8B, 70B, and 405B sizes (text in, text out).

Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Model release date: July 23, 2024

Meta Llama 3.2

The Llama 3.2 collection of multi-lingual large language models is a collection of pre-trained and instruction-tuned generative models in 1B and 3B sizes (text in, text out). There are also quantized versions of these models. The Llama 3.2 instruction-tuned text-only models are optimized for multi-lingual dialogue use cases, including agentic retrieval and summarization tasks. Llama 3.2 1B and 3B models are smaller and less capable derivatives of Llama 3.1.

Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. Note that Llama 3.2 has been trained on a broader collection of languages than these eight officially supported languages. 

Model release date: October 24, 2024

Llama 3.2-Vision

The Llama 3.2-Vision collection of multi-modal large language models is a collection of pre-trained and instruction-tuned image reasoning generative models in 11B and 90B sizes (text and images in, text out). The Llama 3.2-Vision instruction-tuned models are optimized for visual recognition, image reasoning, captioning, and answering general questions about an image.

Supported languages: For text only tasks, English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai are officially supported. Llama 3.2 has been trained on a broader collection of languages than these eight supported languages. Note that for image+text applications, English is the only language supported.

Model release date: September 25, 2024

Meta Llama 3.3

The Meta Llama 3.3 multi-lingual large language model is a pre-trained and instruction-tuned generative model in 70B (text in, text out).

Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Model release date: December 6, 2024

Large language models, including Llama 3.2, are not designed to be deployed in isolation but instead should be deployed as part of an overall AI system with additional safety guardrails as required. Developers are expected to deploy system safeguards when building agentic systems.

Llama 3.3, Llama 3.2 text-only models, and Llama 3.1 support the following built-in tools:

Brave Search: A tool call to perform web searches.

Wolfram Alpha: A tool call to perform complex mathematical calculations.

Code Interpreter: A tool call that enables the model to output python code.

Note the Llama 3.2 vision models don’t support tool calling with text+image inputs.

The Llama Stack

At this point, you may be confused about which of the many Llama models to use for any given application and environment, and how they all fit together. You may find the Llama Stack (see block diagram below) helpful and useful. While Llama Stack emphasizes Llama models, it also provides adapters to related capabilities, such as vector databases for retrieval-augmented generation (RAG).

Llama Stack currently supports SDKs in Python, Swift, Node, and Kotlin. There are a number of distributions you can use, including a local distribution using Ollama, on-device distributions for iOS and Android, distributions for GPUs, and remote-hosted distributions at Together and Fireworks. The general idea is that you can develop locally and then switch to a production end-point easily. You can even run an interactive Llama Stack Playground locally against a remote Llama Stack.

Meta AI’s Llama stack. Meta AI

Running Llama Models

You can run Llama models on Linux, Windows, macOS, and in the cloud. I’ve had good luck running quantized Llama models on a M4 Pro MacBook Pro using Ollama, specifically Llama 3.2 and Llama 3.2-Vision.

It’s worth going through the Llama how-to guides. If you like the LangChain or LlamaIndex frameworks, the Llama integration guides will be helpful.

Llama has evolved beyond a simple language model into a multi-modal AI framework with safety features, code generation, and multi-lingual support. Meta’s ecosystem enables flexible deployment across different platforms, but there are some ongoing legal disputes over training data, and disputes over whether Llama is open source.