Like any major technology company these days, Meta has its own flagship AI’s Genetic Model, called Llama. The blade is somewhat unique among the big models, as it is “open”, which means that developers can download and use it, however, please (with some restrictions). This contradicts models such as Anthropic’s Claude, Google’s Gemini, Xai’s Grok and most of the Openai Chatgpt models, which can only access API.
In the interest of providing developer options, however, Meta has also collaborated with sellers such as AWS, Google Cloud and Microsoft Azure, to make Llama versions available. In addition, the company publishes tools, libraries and recipes in the Llama cookbook to help developers to refine, evaluate and adapt the models to their field. With younger generations such as Llama 3 and Llama 4, these possibilities have been expanded to include inherent multimodal support and wider Rollouts Cloud.
Here is everything you need to know about Meta’s Llama, from its capabilities and versions where you can use it. We will maintain this post up -to -date as post -release upgrades and introduces new Dev tools to support the use of the model.
What is the Llama?
The blade is a family family – not just one. The latest version is Llama 4? It was released in April 2025 and includes three models:
- Detector: 17 billion active parameters, 109 billion total parameters and 10 million brands window.
- Unorthodox: 17 billion active parameters, 400 billion total parameters and a window of environment of 1 million brands.
- Majestic: It has not yet been released, but will have 288 billion active parameters and 2 trillion total parameters.
(In data science, brands are subdivided into pieces of raw data, such as syllables “fan”, “tas” and “tic” in the word “fantastic”)
The framework of a model or environment window refers to the input data (eg text) that the model considers before creating an exit (eg additional text). The long frame can prevent models from “forgetting” the content of recent documents and data and bring out the subject and deviate incorrectly. However, larger environmental windows can also lead to the model by “forgetting” certain safety messages and are more prone to the production of content aligned with the discussion, which has led some users in the direction. wrong thought.
For reference, the window of 10 million frames promised by the Llama 4 Scout is about the text of about 80 medium novels. The Llama 4 Maverick 1 million window is equal to about eight novels.
TechCrunch event
Francisco
|
27-29 October 2025
All Llama 4 models were trained in “large amounts of non -marked text, image and video data” to give them “broad visual understanding” as well as in 200 languages, according to Meta.
The Llama 4 Scout and Maverick are Meta’s first multimodal models. They have been built using a “mixture” architecture (MOE), which reduces computing load and improves efficiency in training and conclusion. Scout, for example, has 16 experts and Maverick has 128 experts.
Llama 4 Behemoth includes 16 experts and Meta refers to it as a teacher for smaller models.
Llama 4 is based on the Llama 3 series, which included 3.1 and 3.2 models widely used for cloud -based applications and development.
What can the Llama do?
Like other AI genetic models, the Lama can perform a number of different auxiliary tasks, such as coding and answering basic mathematical questions, as well as summarizing documents in at least 12 languages (Arabic, English, German, French, Hindi, Indonesian, Italian, Italian, Portuguese, Portuguese, Portuguese Most text-based workloads are thinking of analyzing large files such as PDF and computing sheets-find within its competence and all Llama 4 models support the image and video input.
Llama 4 Scout is designed for larger work flows and mass data analysis. Maverick is a general model that is better in balancing speed and response speed and is suitable for coding, chatbots and technical assistants. And Behemoth is designed for advanced research, distillation of models and STEM duties.
Llama models, including Llama 3.1, can be designed to utilize third -party applications, tools and tools to perform tasks. They are trained to use brave search to answer questions about recent events. Wolfram Alpha’s API for questions related to mathematics and science. and a Python interpreter for validation of code. However, these tools require proper configuration and are not automatically enabled by the frame.
Where can I use llama?
If you are simply looking to chat with Llama, it feeds the Meta Ai Chatbot experience on Facebook Messenger, WhatsApp, Instagram, Oculus and Meta.Ai in 40 countries. Llama printed versions are used in Meta AI experiences in more than 200 countries and territories.
The Llama 4 Scout and Maverick models are available at llama.com and Meta partners, including the AI platform platform. Behemoth is still in training. Building developers with Llama can download, use or perfect the model on most of the popular cloud platforms. Meta claims to have more than 25 partners hosting Llama, including Nvidia, Databricks, Groq, Dell and Snowflake. And while “access” to Meta’s openly available models is not Meta’s business model, the company makes some money through models’ income distribution agreements.
Some of these partners have created additional tools and services over Llama, including tools that allow models to report privately owned data and allow them to run in lower latent periods.
It is important that the Llama license restricts the way developers can develop the model: Application developers with more than 700 million monthly users must seek special permission from META that the company will grant at its discretion.
In May 2025, Meta launched a new program for providing incentives for newly established businesses to adopt Llama models. Llama for newly established businesses gives companies support from Meta’s Llama team and access to possible funding.
Along with Llama, META provides tools that are intended to make the “safer” model to use:
- Lamaa framework of moderation.
- CyberA cyber risk suite in cyberspace.
- Wall of Llama ProtectionA security protector designed to allow the building of secure AI systems.
- Shieldwhich provides support for filtering the conclusions of the unsafe code produced by LLMS.
Llama Guard tries to detect potentially problematic content whether it is powered-whether it is created-from a Lama model, including content related to criminal activity, child exploitation, copyright violations, hatred, self-injury and sexual abuse. This is not clearly a silver sphere, as Meta’s previous instructions allowed Chatbot to participate in sensual and romantic talks with minors and some reports show that they have turned into sexual conversations. Developers can adapt The categories of excluded content and apply the blocks to all languages support Llama.
Like the blade guard, the bumper can block the text intended for the blade, but only the text intended to “attack” the model and behave in unwanted ways. Meta claims that the Lama guard can defend explicit malicious prompts (ie jailbreaks trying to overcome the built -in Llama security filters) in addition to the prompts they contain “input. “The Llama Wall is working to detect and prevent risks such as direct injection, insecure code and dangerous tool interactions.
As far as CybersecVal is concerned, it is less a tool than a collection of reference criteria for measuring model safety. Cyberseceval can evaluate the risk of a Llama model (at least according to META criteria) to application developers and end users in areas such as “automated social engineering” and “escalation of offensive businesses in cyberspace”.
Llama’s restrictions
Llama comes with some dangers and restrictions, like all AI genetic models. For example, while its most recent model has multimodal characteristics, they are mainly limited to English for now.
Zoom in, Meta used a set of pirate e -books and articles to train Llama models. A federal judge recently with Meta in copyright lawsuit dealing with the company by 13 authors of books, ruling that the use of copyright -protected works fell under “fair use”. However, if the Lama overturns a copyright -protected excerpt and one uses it in a product, they could possibly violate copyright and be responsible.
Meta also trains its AI in Instagram and Facebook positions, photos and captions and makes it difficult for users to be excluded.
Planning is another area where it is wise to walk slightly when using Llama. This is due to the fact that the blade can – perhaps more than AI’s genetic counterparts – produce Buggy or unsafe code. On LivecodebenchA reference point that tests AI models in competitive coding problems, Meta’s Llama 4 Maverick model has achieved a 40%rating. This is compared to 85% for GPT-5 high of Openai and 83% for Xai’s Grok 4 quickly.
As always, it is best to have a revision of human experts any code created by AI before incorporating it into a service or software.
Finally, as with other AI models, Llama models are still guilty of creating reasonable but false or misleading information, whether it is coding, legal guidance or emotional conversations with people with AI.
This was originally published on September 8, 2024 and is regularly updated with new information.
