Meta Llama: All you need to know about the open artificial intelligence model

Like every large technical company these days, Meta has a model of pioneering artificial intelligence, called Llama. Llama is somewhat unique between the main models in that it is “open”, which means that developers can download and use as they like (with some restrictions). This is unlike models such as Clauds, Jimini from Google, Xai’s Grok, and most Chatgpt models of Openai, which can only be accessed through application programming facades.

However, it is in the interest of giving developers, Meta is also involved with sellers, including AWS, Google Cloud, and Microsoft Azure, to allow llama versions hosted by the cloud. In addition, the company publishes tools, libraries and recipes in the Llama cooking book to help developers control, evaluate and adapt models to their field. With newer generations like Llama 3 and Llama 4, these capabilities expanded to include original multimedia support and wider clouds.

Here is everything you need to know about Llama’s Meta, from its capabilities and publications to the place you can use. We will keep this post update with the version of Meta updates and offers new DeV tools to support the use of the form.

What is Lama?

Lama is a family of models – not just one. The latest version Lama 4; It was released in April 2025 and includes three models:

• searchlight: 17 billion active teachers, 109 billion teachers, and a context window of 10 million symbols.

• Maverick: 17 billion active teachers, 400 billion teachers, and a context window of one million symbols.

• giant: It has not been released yet, but it will have 288 billion active teachers and 2 trillion parameters.

(In data science, distinctive symbols of raw data are organized, such as “fan”, “tas” and “tic” in the word “wonderful”.

The context of the model, or context window, indicates the input data (for example, the text) that the model sees before the creation of the output (for example, additional text). The long context can prevent models from “forgetting” the content of modern documents and data, from deviating from the subject and extrapolation of the wrong. However, the longest context windows can also It leads to a “forget” model Be more likely to produce content that is in line with the conversation, which led to some users towards that Fake thinking.

To refer to it, the 10 million context window that Llama 4 Scout is equal to almost the text of about 80 medium novels. Lama 4 mafrickA context window of one million people equal to about eight novels.

All Llama 4 models have been trained on “large quantities of texts, image and unacceptable video” to give them “wide visual understanding”, as well as 200 languages, according to Meta.

Llama 4 Scout and MAVERICK is the first multimedia mode in Meta. It has been designed using the “Mix Experts” structure (MEE), which reduces arithmetic load and improves training and inference. Scout, for example, has 16 experts, and MAVERICK has 128 experts.

Llama 4 Beheemoth includes 16 experts, and Meta refers to this as a teacher of smaller models.

Llama 4 depends on the Llama 3 series, which included a wide 3.1 and 3.2 models used for applications that were widely seized and published the cloud.

What can Lama do?

Like other artificial intelligence models, Llama can perform a range of different auxiliary tasks, such as coding and answering basic mathematics questions, as well as summarizing documents in at least 12 languages ​​(Arabic, English, German, French, Indian, Indonesian, Italian, Portuguese, Hindu, Spanish, Tagalog, Thai, and Vitami). Most of the burden of text-based work-think about analyzing large files such as PDFS and data tables-is within its jurisdiction, support all Llama 4 texts, a picture and a video insert.

Llama 4 Scout is designed for the longest workflow and a huge data analysis. MAVERICK is a better general model in the balance of thinking power and response speed and is suitable for coding, chat chat, and technical assistants. The giant has been designed for advanced research, the form of the form, and STEM tasks.

Llama models, including Llama 3.1, can be formed, to take advantage of third -party applications, tools and a application programming interface to perform tasks. They were trained to use brave search for answering questions about recent events; Wolfram Alpha API for mathematics and science inquiries; Bethon translator to verify the validity of the code. However, these tools require appropriate composition and are not enabled automatically outside the box.

Where can I use Lama?

If you are looking to chat with Llama simply, this is it Meta Ai Chatbot On Facebook Messenger, WhatsApp, Instagram, Oculus and Meta.ai in 40 countries. Llama versions that were seized in Meta AI experiences are used in more than 200 countries and regions.

Models Llama 4 Models Scout and MAVERICK are available on Llama.com and Meta partners, including the platform of artificial intelligence developers. The giant is still in training. The developers with Llama can download, use, or adjust the model through most of the famous cloud platforms. Meta claims that she has more than 25 partners hosting Llama, including NVIDIA, Databrics, Groq, Dell and Snowflake. Although the “sale of access” to the available models is not the Meta business model, the company makes some money through Revenue sharing agreements With host hosts.

Some of these partners have built additional tools and services at the head of Llama, including tools that allow models to refer to ownership data and enable them to operate in low time time.

More importantly, Lama license It restricts how developers can publish the formApplication developers who have more than 700 million users per month request a special license from Meta that the company will give it at its estimate.

In May 2025, she launched a dead New program To motivate startups to adopt Llama models. Llama provides startups for companies support from the Llama team in Meta and access to potential financing.

Besides Llama, Meta provides tools aimed at making the “safer” model for use:

• Lama goalkeeperModeration framework.

• CYBERSECEVALCyber ​​security risk evaluation wing.

• Lama protection wallA security study designed to enable the building of safe artificial intelligence systems.

• CodeWhich provides support to liquidate the time of inference to the unsafe symbol produced by LLMS.

Llama Guard tries to discover content that is likely to be in or created-through the Llama model, including the content related to criminal activity, children’s exploitation, copyright violation, hatred, self-harm and sexual assault.

However, it is clear that it is not a silver bullet since then Previous private guidelines Chatbot was allowed to engage in sensory and romantic chats with minors, and some reports show that they turned to Sexual talks. Developers can Assignment The categories of prohibited content and the application of blocks to all languages ​​supported by Llama.

Like Llama Guard, PROPER Guard can prevent the intended text for Llama, but only the text aims to “attack” the model and obtain it to act in unwanted ways. Meta claims that the Llama goalkeeper can defend explicit harmful claims (that is, prison operations trying to wander around Lama’s safety filters) in addition to claims that contain “Included injecting“Llama’s protection wall detects and prevents risk such as instant injection, unsafe code, and risk -framed tools.

For Cyberseceval, it is a tool less than a set of criteria for measuring the security of the model. Cyberseceval can evaluate the risks offered by the Llama model (at least according to META standards) for applicants and final users in areas such as “automatic social engineering” and “offensive electronic scaling operations”.

Lama restrictions

Lama comes with some risks and restrictions, like all artificial intelligence models. For example, although its latest model has multimedia features, it is mainly limited to English at the present time.

Zoom, Meta use a collection of data from pirated e -books And articles to train Llama models. A federal judge recently sided Meta in a lawsuit for copyrights filed against the company by 13 writers of the book, and decided that the use of copyright works decreases under “fair use”. However, if Lama Extension An excerpt protected by copyright and someone who uses it in a product, can violate copyright and be responsible.

Meta also trains artificial intelligence on Instagram, Facebook, photos and labels, and It makes it difficult for users to cancel the subscription.

Programming is another field where it is wisely to walk lightly when using Llama. That is because Lama may – perhaps more than her counterparts from artificial intelligence – Production of a vehicle that is drawn by animals or insecure. on LiveCOOOOOBENCH, standard This tests artificial intelligence models on competitive coding problems, the Meta’s Llama 4 MAVERICK model has achieved 40 % score. This is compared to 85 % for Openai GPT-5 High and 83 % for Xai’s Grok 4 Fast.

As always, it is better to have human experts to review any symbol created by artificial intelligence before it is integrated into a service or program.

Finally, as with other Amnesty International models, Llama models are still guilty of generating reasonable but wrong or misleading information, whether in coding, legal guidance or Emotional conversations with artificial intelligence characters.

This was originally published on September 8, 2024, and it is regularly updated with new information.