Llama 4

Definition

Llama 4 is Meta's next-generation open-weight model family, released in 2025, introducing a mixture-of-experts architecture and native multimodality across a range of model sizes including Scout, Maverick, and the largest Behemoth model.

Overview

Llama 4 marked a significant architectural shift from Llama 3: instead of a single dense network, Llama 4 models use a mixture-of-experts design, where only a subset of the model's total parameters are activated for any given input, allowing for a much larger total parameter count while keeping inference cost more manageable. Meta released the family under names such as Llama 4 Scout and Llama 4 Maverick, with a much larger Llama 4 Behemoth model referenced as a still-larger, more capable sibling. Llama 4 was also designed as natively multimodal from pretraining, handling text and images together rather than adding vision capability afterward, and Scout in particular was notable for an unusually long context window aimed at processing very large documents or codebases in a single pass. As with prior Llama generations, Llama 4 weights are released for download and self-hosting under Meta's community license, continuing to anchor a large ecosystem of open-weight derivatives, fine-tunes, and research built on top of it, and competing with other open and semi-open model families such as Qwen, DeepSeek-V3, and Mistral models.

Key Features

• Mixture-of-experts architecture activating only a subset of parameters per input
• Released as multiple variants: Scout, Maverick, and the larger Behemoth
• Natively multimodal, trained on text and images together from the start
• Notably long context window in the Scout variant
• Open-weight release under Meta's community license
• Positioned to compete with other open model families like Qwen and DeepSeek

Use Cases

Frequently Asked Questions

Related in AI Models