Meta unveils Llama 4, its new open-source AI

The parent company of Facebook, Instagram, and WhatsApp unveiled its new family of AI models, Llama 4, open source and featuring multimodal algorithms

Meta announced the launch of its new family of artificial intelligence models, called Llama 4, based on multimodal algorithms and capabilities comparable to those of OpenAI’s GPT-4o, Google’s Gemini 2.0 Flash, and DeepSeek-V3.

The parent company of social media platforms Instagram, Facebook, and WhatsApp has stated that its new functionality is open source, although it has been criticized. The truth is that “these new algorithms stand out for being multimodal and having large-scale context windows, ranging from 1 million to 2 trillion tokens.”

The first model presented by Meta is Llama 4 Scout, described as “the world’s best multimodal model in its category and outperforms all previous generations of Llama.”

Its architecture integrates 17 billion active parameters, 10 million tokens, and a system based on a mixture of experts, “which integrates 16 submodels specialized in various tasks, topics, and multimedia formats that are activated based on user requests,” to outperform Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 in “deep data analysis, extensive code reviews, generation of complex summaries, and legal evaluation of large documents.”

Then there is Llama 4 Maverick, trained simultaneously with text, images, and video without the need for prior labeling. It incorporates 128 experts, allowing it to reach a total of 400 billion parameters, keeping 17 billion active per token to compete with DeepSeek V3 and surpass GPT-40, Gemini 2.0, and Claude 3 Opus in reasoning, coding, and understanding visual and textual content.

The first two AI models are followed by Llama 4 Behemoth, “the company’s most advanced model and one of the smartest large language models in the world.” It is in development and not publicly available; it has 288 billion active parameters and nearly 2 trillion in total. It is used to improve the quality of smaller systems through co-distillation techniques. Tests showed that Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in multiple assessments related to STEM (science, technology, engineering, and mathematics) disciplines.

It’s important to note that Llama 4 Maverick and Scout are now available for download on the official Meta website and the Hugging Face platform. They are expected to soon be integrated into major cloud infrastructures, edge computing solutions, and services offered by global integrators.

M.Pino

Source: wired

(Reference image: Mohamed Nohassi on Unsplash)

Visit our news channel on Google News and follow us to get accurate, interesting information and stay up to date with everything. You can also see our daily content on X/Twitter and Instagram

You might also like