Meta’s Llama 4 Models Denote a New Era of AI Intelligence

Meta is unveiling the first suite of models in the Llama 4 “herd,” marking the beginning of a new, intelligent era for the Llama ecosystem, according to the company. These models—Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth—each address unique needs in the AI space—particularly around contextual windows and model intelligence.

The first of these models, Llama 4 Scout, is a general-purpose, 17 billion active parameter model with 16 experts that delivers state-of-the-art performance for its class, according to Meta. At its core, this model dramatically increases supported context length, from 128K in Llama 3 to 10 million tokens—which corresponds to text longer than 5M words. This unlocks a variety of new opportunities—such as multi-document summarization, analyzing extensive user activity, and reasoning over large codebases.

Llama 4 Maverick, the larger of the two Llama 4 models, is a 17 billion active parameter model with 128 experts. Designed with industry-leading performance in image and text understanding, Llama 4 Maverick is a multimodal model that excels at precise image understanding and creative writing. Llama 4 Maverick exceeds similar models such as GPT-4o and Gemini 2.0 on coding, reasoning, multilingual, long-context, and image benchmarks, and is competitive with the much larger DeepSeek v3.1 on coding and reasoning, according to Meta.

Both Llama 4 Scout and Llama 4 Maverick are designed to offer native multimodal intelligence at a compelling price while outperforming significantly larger models, Meta explained. Maintaining a commitment to openness, Llama 4 Scout and Llama 4 Maverick are available for download on llama.com and Hugging Face.

Llama 4 Behemoth—which is presently only available in preview—is a 288 billion active parameter teacher model with 16 experts. As one of the world’s most powerful and smartest LLMs, Llama 4 Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks, according to Meta. Teaching the smaller Llama 4 models, Llama 4 Behemoth is a multimodal mixture-of-experts model that delivered significant improvements across end-task evaluation metrics for Llama 4 Maverick.

Regarding safety, the Llama 4 suite of models adheres to Meta’s best practices outlined in its Developer Use Guide: AI Protections, instilling mitigations at each layer of model development.

To learn more about Meta’s Llama 4 models, please visit https://ai.meta.com/.