Meta introduced Scout, Maverick and Behemoth, the three models that make up the Llama 4 range, designed both for developers and to enrich the capabilities of Meta AI.
Has Meta closed the gap in the AI race? This Saturday, April 5, 2025, the California-based group unveiled the outlines of Llama 4, a new generation of artificial intelligence models that, according to its internal tests, would far outperform previous iterations and could shake up the most powerful technologies on the market, starting with those from Google and OpenAI.
Scout, Maverick and Behemoth, the three new flagship models from Meta
In a blog post, Meta detailed the capabilities of its three new flagship models, “which combine multimodal intelligence with an attractive price” :
- Llama 4 Scout: Billed as a “general-purpose model,” Llama 4 Scout stands out for its low resource consumption while integrating more than 17 billion parameters. According to Meta, it can handle up to 10 million tokens, which would make it an effective model for summarizing multiple documents, reasoning from “large code bases,” or analyzing user activity.
- Llama 4 Maverick: Considered a “general-purpose model,” Llama 4 Maverick also includes 17 billion parameters and, according to Meta, is particularly well-suited for assistance tasks or creative writing. It is said to outperform GPT-4o or Gemini 2.0 in these areas, as well as in image understanding. Within days, Llama 4 Maverick jumped straight to second place in the Chatbot Arena, a ranking that objectively evaluates the models available on the market.
- Llama 4 Behemoth: Available in preview, Llama 4 Behemoth is the most powerful model in the range. With 228 billion parameters, it is said to deliver high performance in image analysis and mathematical problem solving, according to Meta. However, it has not yet been deployed.
Note that none of these models adopt an “advanced reasoning” approach, which consists of taking time to reflect on and analyze information before generating a response, unlike the Gemini 2.5 models, unveiled a few days ago by Google, or the “o” series from OpenAI.
Models gradually integrated into Meta AI… but unavailable in Europe
According to Meta, the models will gradually be integrated into Meta AI, the chatbot available on WhatsApp, Instagram, Facebook and Messenger, which was recently launched in Europe. According to the specialist media TechCrunch, Meta AI is using these models in 40 countries.
But in Europe, you’ll probably have to wait. While Llama 4 Scout and Llama 4 Maverick are available on the Llama.com platform and on Hugging Face, they don’t appear to be accessible to European developers. On a dedicated page, Meta specifies that the rights to use Llama 4 models are “not granted if you are an individual domiciled in the European Union or a company whose principal place of business is in the European Union.”
To date, no availability date in the European Union has been announced.
