Like all large technological companies these days, Meta has its own flagship model of generative AI, called Llama. Llama is somewhat unique among the main models insofar as it is “open”, which means that developers can download it and use it as they see fit (with certain limitations). This contrasts with models like Claude d’Anthropic, Gemini de Google, Xai Grok and most of the Openai Chatgpt models, which are only accessible via APIs.
However, in order to give the developers the choice, Meta also teamed up with suppliers, including AWS, Google Cloud and Microsoft Azure, to make available versions hosted in the Llama Cloud. In addition, the company publishes tools, libraries and recipes in its LLAMA recipe book to help developers refine, assess and adapt the models to their field. With new generations like Llama 3 and Llama 4, these features have extended to include a native multimodal management and wider cloud deployments.
Here is everything you need to know about Meta’s Llama, its capacities and editions to the place where you can use it. We will keep this article up to date as Meta publishes upgrades and will introduce new development tools to support the use of the model.
What is Lama?
Llama is a family of models, not just one. The latest version is Llama 4; It was released in April 2025 and includes three models:
- Scout: 17 billion active parameters, 109 billion parameters in total and a contextual window of 10 million tokens.
- Non-conformist: 17 billion active parameters, 400 billion parameters in total and a contextual window of 1 million tokens.
- Monster: Not yet published but will have 288 billion active parameters and 2,000 billion parameters in total.
(In data science, tokens are pieces of raw data subdivided, such as the syllables “Fan”, “TAS” and “TIC” in the word “fantastic”.)
The context of a model, or contextual window, refers to the input data (for example, the text) that the model takes into account before generating an output (for example, additional text). A long context can prevent “forgetting” models from the content of recent documents and data, as well as deviating from the subject and wrongly extrapolating. However, longer contextual windows can also bring the model to “forget” certain security railings and be more inclined to produce content in line with the conversation, which led certain users to a delirious thought.
For reference, the context window of 10 million promised by Llama 4 Scout is roughly the text of around 80 novels on average. The contextual window of one million Llama 4 Maverick is equivalent to about eight novels.
Techcrunch event
San Francisco
|
October 27-29, 2025
All Llama 4 models have been formed on “large amounts of text, image and video data not labeled” to give them “a large visual understanding”, as well as on 200 languages, according to Meta.
Llama 4 Scout and Maverick are the first natively multimodal models with open weights of Meta. They are built using an architecture “mixture of experts” (MOE), which reduces the calculation load and improves the efficiency of training and inference. Scout, for example, has 16 experts and Maverick 128 experts.
Llama 4 Behemoth includes 16 experts, and Meta refers to it as a teacher for small models.
Llama 4 is based on the Llama 3 series, which included 3.1 and 3.2 models widely used for instructions oriented applications and the cloud deployment.
What can the lama do?
Like other generative AI models, Llama can perform a range of different assistance tasks, such as code and answer basic mathematical questions, as well as summarize documents in at least 12 languages (Arabic, English, German, French, Hindi, Indonesian, Italian, Portuguese, Hindi, Spanish, Tagalog, Thai and Vietnamese). Most of the work -based workloads (think of the analysis of large files such as PDFs and calculation sheets) fall under its skill, and all LLAMA 4 models support text, image and video entry.
LLAMA 4 Scout is designed for longer workflows and a massive data analysis. Maverick is a general model that manages to balance reasoning power and response speed and suitable for coding, chatbots and technical assistants. And Behemoth is designed for advanced research, distillation of models and STEM tasks.
LLAMA models, including LLAMA 3.1, can be configured to use applications, tools and third -party APIs to perform tasks. They are trained to use Brave Search to answer questions about recent events; The Wolfram Alpha API for requests related to mathematics and sciences; and a Python interpreter to validate the code. However, these tools require an appropriate configuration and are not automatically activated from the start.
Where can I use Lama?
If you are just looking to chat with Llama, he feeds the Meta Ai chatbot experience on Facebook Messenger, Whatsapp, Instagram, Oculus and Meta.ai in 40 countries. Llama refined versions are used in Meta AI experiences in more than 200 countries and territories.
The Llama 4 Scout and Maverick models are available on llama.com and Meta partners, including the development platform of IA Hugging Face. Behemoth is still in training. Developers who build with Llama can download, use or refine the model to most popular cloud platforms. Meta claims to have more than 25 partners hosting Llama, including Nvidia, Databricks, Groq, Dell and Snowflake. And although “selling access” to the openly available models of Meta is not Meta’s business model, the company earns money thanks to income sharing agreements with model hosts.
Some of these partners have created additional tools and services on LLAMA, in particular tools that allow models to reference proprietary data and to operate with lower latencies.
It is important to note that the Llama license limits the way in which developers can deploy the model: application developers with more than 700 million monthly users must request a special license in Meta that the company will grant its discretion.
In May 2025, Meta launched a new program to encourage startups to adopt its Llama models. Llama for Startups offers companies support the Meta Llama team and access to potential funding.
Alongside Llama, Meta provides tools intended to make the “safer” model to use:
- Lama GuardA moderation framework.
- CyberscevalA sequence of cybersecurity risk assessment.
- Lama firewallA safety railings designed to create secure AI systems.
- Coded shieldwhich supports filtering at the time of the inference of the unsecured code produced by the LLM.
Llama Guard is trying to detect potentially problematic content supplied or generated by an LLAMA model, including content related to criminal activities, children’s exploitation, copyright, hatred, self -harm and sexual abuse.
That said, it is clearly not a miracle solution since the previous directives of Meta authorized the chatbot to engage in sensual and romantic conversations with minors, and certain reports show that these have turned into sexual conversations. Developers can personalize blocked content categories and apply the blocks to all languages supported by Llama.
Like Llama Guard, Prompt Guard can block the text intended for Llama, but only the text intended to “attack” the model and to make it behave undesirablely. Meta claims that Llama Guard can defend herself against explicitly malicious guests (that is to say the jailbreaks who try to bypass the integrated safety filters of Llama) in addition to the prompts containing “injected entries”. The Llama firewall works to detect and prevent risks such as rapid injection, unsecured code and risky interactions with tools. And Code Shield helps to mitigate the suggestions of unsecured code and offers a secure order execution for seven programming languages.
As for Cybersceval, it is less a tool than a set of benchmarks to measure the safety of models. Cybersceval can assess the risk that a Llama model poses (at least according to META criteria) for application developers and end users in areas such as “automated social engineering” and “the scaling of offensive cyber operations”.
The limits of the lama
Llama has certain risks and limits, such as all generative AI models. For example, although its most recent model has multimodal functionalities, these are currently mainly limited to the English language.
By making a rear zoom, Meta used a set of electronic books and hacked items to form its LLAMA models. A federal judge recently took sides for META in a trial in copyright brought against the company by 13 authors of books, ruling that the use of works protected by copyright for training purposes was “fair use”. However, if there regurgite a copyright protected extract and someone uses it in a product, he could potentially break copyright and be responsible for it.
Meta also controversially leads to her AI on Instagram and Facebook publications, photos and legends, and makes user unsubscribe difficult.
Programming is another area in which it is wise to show lightness when you use Llama. Indeed, Llama could-perhaps more than its generative AI counterparts-produce bogued or unsecured code. On LiveCodebench, a benchmark that tests AI models on competitive coding problems, the Meta Llama 4 Maverick model obtained a 40 %score. This is compared to 85 % for the OPENAI High GPT-5 and 83 % for Grok 4 Fast from XAI.
As always, it is best to ask a human expert to examine any code generated by AI before incorporating it into a service or software.
Finally, as for other AI models, Llama models are always guilty of generating plausible but false or misleading information, whether coding, legal advice or emotional conversations with IA characters.
This was initially published on September 8, 2024 and is regularly updated with new information.
