Welcome Llama 4 Maverick & Scout on Hugging Face

Hugging Face Blog · 1 year ago

Meta released two new large language models, Llama 4 Maverick and Llama 4 Scout, on Hugging Face with native multimodal capabilities and Mixture of Experts architecture. Maverick has 400 billion total parameters with 128 experts (17 billion active), while Scout has 109 billion total parameters with 16 experts (17 billion active), both trained on 40 trillion tokens across 200 languages. The models support context lengths of 1 million tokens for Maverick and 10 million tokens for Scout in their instruction-tuned variants, with immediate integration into Hugging Face transformers and Text Generation Inference for deployment.