Llama 4 Maverick 17B Instruct

Llama 4 Maverick represents Meta's push into efficient, high-capacity multimodal AI through a mixture-of-experts architecture that activates only 17 billion parameters per forward pass while routing across 128 expert networks for a total of 400 billion parameters. This design intent focuses on delivering strong performance with improved throughput compared to dense models of equivalent total size. The model employs early fusion to natively process text and images together from the ground up, supporting multilingual interaction across 12 languages for both input and code generation. The architecture is optimized for vision-language tasks, instruction-following, and general-purpose multimodal reasoning, making it well-suited for applications that require understanding and reasoning about visual content alongside natural language.

The model was trained on a curated mixture of public, licensed, and Meta-platform data encompassing approximately 22 trillion tokens, with instruction-tuning applied to cultivate assistant-like behavior, image reasoning capabilities, and general-purpose interaction. This training lineage combines breadth of data sources with targeted post-training to develop the model's practical strengths in multimodal understanding. Released under the Llama 4 Community License with open weights, Maverick provides accessibility for both research exploration and commercial deployment where advanced multimodal capabilities and efficient inference through its mixture-of-experts design are priorities.