Llama 4 Scout is a 109 billion parameter Mixture-of-Experts (MoE) multimodal LLM with 17 billion active parameters and 16 experts, designed for efficient, large-scale reasoning over both text and images. It excels in tasks requiring extremely long context—supporting up to a 10 million token window—making it particularly strong for multi-document summarization, codebase analysis, and personalized automation based on extensive user data[1][2][4][6].

Some other noteworthy features of Llama 4 Scout include native support for both text and image inputs (multimodality), multilingual capabilities across twelve major languages, and compute efficiency that allows it to run on a single H100 GPU with quantization. Its architecture enables high performance in image understanding (e.g., visual question answering), long-context retrieval tasks (“needle-in-a-haystack”), coding benchmarks, and advanced reasoning—all while maintaining lower inference costs compared to similarly sized dense models.

Metric	Value
Parameter Count	109 billion total / 17 billion active
Mixture of Experts	Yes
Active Parameter Count	17 billion
Context Length	Up to 10 million tokens (currently limited to ~300K–327K on Together.ai)
Multilingual	Yes
Quantized*	No

*Quantization is specific to the inference provider and the model may be offered with different quantization levels by other providers.