Meta: Llama 4 Maverick

Llama 4 Maverick 17B Instruct (128E) is a cutting-edge, high-capacity multimodal language model developed by Meta. Built on a sophisticated mixture-of-experts (MoE) architecture, it leverages 128 experts and 17 billion active parameters per forward pass (400B total) to deliver exceptional performance. This model excels in vision-language tasks, supporting multilingual text and image input while generating multilingual text and code output across 12 supported languages. Maverick is instruction-tuned for assistant-like behavior, advanced image reasoning, and general-purpose multimodal interaction, making it highly versatile. Key features include early fusion for native multimodality and an impressive 1 million token context window, allowing for deep contextual understanding. Trained on a vast dataset of approximately 22 trillion tokens, comprising public, licensed, and Meta-platform data, its knowledge cutoff is August 2024. Released on April 5, 2025, under the Llama 4 Community License, Maverick is suitable for both research and commercial applications demanding advanced multimodal understanding and high model throughput. It offers capabilities such as vision, functions, code, and streaming, with a maximum output of 4K tokens. Pricing is set at $0.15 per 1M input tokens and $0.60 per 1M output tokens, available in the STARTER access tier.