Meta Releases Llama 4 Scout and Maverick with Native Multimodal Support

Meta has officially released the Llama 4 model family, introducing two distinct variants designed for different use cases. Llama 4 Scout is a 17-billion active parameter model optimized for efficiency and edge deployment, while Llama 4 Maverick employs a Mixture-of-Experts architecture with 17 billion active parameters from a 400 billion total parameter pool.

Both models feature native multimodal capabilities, processing images and videos alongside text without requiring separate vision encoders. This integration enables more coherent reasoning across modalities and reduces the architectural complexity for developers building multimodal applications.

Llama 4 Scout achieves competitive performance with GPT-4o on several benchmarks while requiring significantly less computational resources. Maverick, targeting research and enterprise applications, matches or exceeds GPT-4o performance on most standard benchmarks.

The models are available under Meta's custom Llama license, which permits commercial use for companies with fewer than 700 million monthly active users. The release continues Meta's strategy of maintaining competitive open-weight models to drive ecosystem adoption.