Meta releases Llama 4 Scout and Maverick as first open-weight MoE multimodal models

Meta has announced the release of Llama 4 Scout and Llama 4 Maverick, introducing the first open-weight natively multimodal models built with a mixture-of-experts (MoE) architecture. These models are designed to support the entire Llama ecosystem with unprecedented context length support and efficient deployment capabilities.

• Llama 4 Scout features 17 billion active parameters with 16 experts, fitting on a single H100 GPU with Int4 quantization and supporting a 10M input context length.
• Llama 4 Maverick uses 17 billion active parameters with 128 routed experts plus a shared expert, allowing it to run on a single H100 DGX host while lowering inference latency.
• Both models utilize early fusion to seamlessly integrate text and vision tokens, trained on over 30 trillion tokens including diverse text, image, and video datasets.
• Meta also previewed Llama 4 Behemoth, a teacher model that outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks like MATH-500 and GPQA Diamond.

The release aims to provide developers with efficient, high-quality multimodal intelligence for building next-generation consumer and business applications.