Thought Leadership
Meta unveiled the latest generation of its large language models (LLM) at the weekend. With Llama 4 Scout and Llama 4 Maverick, the company is presenting what it describes as its “most advanced models yet.”
Both models are part of the company’s broader strategy to develop open source AI tools. According to Meta, the new language models can process and generate different types of data – including text, images, video and audio. This makes them fully multimodal systems. “These Llama 4 models mark the beginning of a new era for the Llama ecosystem,” Meta wrote in a blog post. “This is just the beginning for the Llama 4 collection.”
Llama 4 Scout has a context window of 10 million tokens. Meta claims that Scout can outperform Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as open-source competitor Mistral 3.1 “on a wide range of commonly reported benchmarks”. Despite its performance, the model is said to be compact enough to run on a single Nvidia H100 GPU.
The larger Llama 4 Maverick model was developed to compete with OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash. Meta claims that Maverick delivers comparable results to DeepSeek-V3 for programming and reasoning tasks, but requires “less than half the active parameters”.
Llama 4 “Behemoth” as a teacher model
In addition, Meta announced a preview model called Llama 4 Behemoth, described as “one of the world’s most intelligent LLMs and our most powerful model to date”. According to the company, Behemoth will serve as a “teacher” for the new models.
The continuously trained Behemoth model will contain a total of 2 trillion parameters, 288 billion of which will be active at all times. The model, which uses a “Mixture of Experts” (MoE) architecture to optimize resource usage, is expected to outperform GPT-4.5 and Claude Sonnet 3.7 “on multiple STEM benchmarks”.
Meta CEO Mark Zuckerberg describes Behemoth as “the most powerful base model in the world”. The company plans to present further technical details and product announcements during its upcoming developer event LlamaCon on April 29.
The development of Llama 4 was not without its difficulties. According to a report by The Information, Meta delayed the release due to initial performance problems with logical thinking and mathematical tasks. There were also concerns that the models were less efficient at generating human-like language conversations than those of OpenAI.
Scout and Maverick are available on Llama.com and from Meta’s partners, including AI development platform Hugging Face, while Behemoth is still in development. Meta says Meta AI, its AI-powered assistant for apps such as WhatsApp, Messenger and Instagram, has been updated to use Llama 4 in 40 countries. Multimodal features are limited to the US in English for now.
