- Meta launches Llama 4 with three variants: Scout, Maverick, and Behemoth, with a multimodal approach and open source.
- The models incorporate Mixture of Experts (MoE) architecture and context windows of up to 10 million tokens.
- Scout and Maverick are already available on apps like WhatsApp and Instagram, while Behemoth is still in development.
- Distribution of Llama 4 is restricted outside the US by legal frameworks, particularly in the EU.
Meta has taken another step in the evolution of its Artificial Intelligence ecosystem with the launch of Llama 4 , a collection of models that aim to compete with current industry giants like GPT-4 and Gemini. With a revamped architecture and the addition of multimodal capabilities , Mark Zuckerberg's company is trying to strengthen its position in a race characterized by innovation and rapid development.
The Llama 4 family arrives with three main models: Scout, Maverick, and Behemoth , each with distinct features designed to cover a wide range of needs, from everyday tasks to highly specialized uses. Furthermore, Meta's commitment remains clear: maintaining open source as a differentiating factor, which positions these models prominently within the landscape of accessible tools for businesses and individual developers. For the latter, Llama 4's open source code offers the opportunity to experiment and develop custom applications.
A new architecture for a new generation
One of the most significant aspects of Llama 4 is its architecture based on the Mixture of Experts (MoE) technique . This approach allows only the necessary parts of the model to be activated for a given task, rather than activating all its parameters in every query. This not only improves efficiency but also reduces computational load and speeds up responses. With this architectural advancement, AI developers can explore new possibilities, such as those presented in the article on Microsoft Loop.
Llama 4 Scout and Maverick share a base of 17.000 billion active parameters , although they differ in the number of experts they use: Scout uses 16, while Maverick employs 128. This configuration differentiates their capabilities, focusing Scout on technical tasks such as code analysis or documentation synthesis , while Maverick focuses on more general functions such as creative writing or text and image interpretation . For those interested in code analysis, Llama 4's architecture can be fundamental, and you can find more information in the post about Python.
In contrast, Behemoth represents a push towards the cutting edge of model development . Although still in the training phase, it has been revealed that it boasts 288.000 billion active parameters distributed among 16 experts. According to Meta, its initial tests place it above competitors such as GPT-4.5, Claude 3.7, and Gemini 2.0 Pro in benchmarks focused on STEM fields.
Multimodality as the key to the future
Multimodality is another hallmark of this new generation . Scout and Maverick are designed to understand and work with different types of input, such as text , images , and even video , thus enabling the development of more versatile and adaptable assistants. This translates into more sophisticated applications within corporate, educational, or creative environments, where the simultaneous interpretation of visual and textual information is essential. In this context, the integration of models like Llama 4 could contribute to improving the capabilities of modern tools, similar to what is seen in Google Colab.
One of the most innovative features is the introduction of context windows of up to 10 million tokens . With this capability, the models can handle extremely large blocks of data, facilitating in-depth analysis of extensive documents or complex codebases. This expansion of context is unprecedented in previous models and could become a standard for future AI releases globally.
Thanks to the MoE architecture, Llama 4 models also consume fewer resources during training and inference , which is essential for scaling their use without requiring excessively expensive infrastructure. Meta mentions that Scout can run on a single Nvidia H100 GPU , democratizing its accessibility for small businesses or developers with more limited resources. This resource-efficient approach is similar to that of Wi-Fi connection systems , where resource optimization is key to tangible benefits.
Availability and international deployment
The Scout and Maverick models are now available on several Meta platforms , including WhatsApp, Instagram, Facebook Messenger, and the Meta AI website. For those seeking more technical access or wishing to develop them independently, versions are available on llama.com and the popular Hugging Face platform . This international availability marks a significant step forward in accessing Artificial Intelligence, as discussed in the article on Microsoft's latest news.
However, not everyone will have equal access to these advancements. Meta has confirmed that both multimodal functions and model distribution are restricted in the European Union for legal reasons related to data protection and artificial intelligence regulations. This means that companies located within the EU cannot, for now, use or redistribute them without a specific license.
Furthermore, corporations with more than 700 million monthly active users will need to apply for additional permits to implement these technologies. This could slow their rollout in markets like Europe, although Meta has indicated its intention to address these obstacles in future updates.
One step closer to the reasoning model
Beyond its current models, Meta has already publicly announced that it is working on a fourth variant: Llama 4 Reasoning . As its name suggests, it will focus on improving logical reasoning, something its current versions do not yet fully integrate. The introduction of this model would mark Meta's first attempt to compete in an area where OpenAI and other companies have already made significant strides.
Currently, Llama 4 models respond more freely to social or political issues, something Meta has adjusted so that new versions are less reticent about addressing controversial topics. This reduction in filters follows the trend seen in recent models like Grok 3 and is part of a broader shift in the industry toward more open and less censored interactions.
Furthermore, Meta has hinted that Llama 4 Behemoth could serve as a basis for creating smaller, more specialized models , allowing its power to be applied in more specific contexts without the need to deploy the entire infrastructure.
The new Llama 4 generation positions Meta firmly within the competition for the development of advanced AI models . With a combination of open source, efficient architecture, and multimodal capabilities, Scout and Maverick already offer solutions applicable today. Meanwhile, Behemoth promises to be a cornerstone upon which to build even more powerful models in the future. Although its distribution is limited by legal issues in some regions, the technical advancement appears undeniable and reinforces Meta's commitment to leading the open artificial intelligence sector.
