Llama 4 Scout

What it is

Llama 4 Scout is part of the Llama 4 herd, which Meta released on 5 April 2025. The herd was Meta's first set of mixture-of-experts models and its first natively multimodal Llama models. Natively multimodal means text and vision are handled in one model from the start, using early fusion of text and vision tokens, rather than by attaching a vision component afterwards.^[1]

A mixture-of-experts model splits its network into many specialist parts, called experts, and routes each token to a few of them. Only the active parameters are used for a given token, so a large model can be cheaper to run than its total size suggests.

Origins and training

Scout has 17 billion active parameters, 16 experts and 109 billion total parameters. Meta says it fits on a single NVIDIA H100 GPU when quantised to Int4.^[1][2]

It was trained on about 40 trillion tokens of text, image and video data, with an August 2024 knowledge cutoff. Training used 5.0 million GPU hours.^[2]

Main features

Context window. Scout's context window is 10 million tokens. A context window is the amount of text, measured in tokens, that the model can take into account at once. Llama 3 had 128K. Scout was pretrained and post-trained at 256K context.^[1][2]

Architecture. Scout uses an iRoPE design, which means interleaved attention layers without positional embeddings. Its vision encoder is based on MetaCLIP.^[1]

Benchmark results. On instruction-tuned benchmarks, Scout scored 74.3 on MMLU Pro, 57.2 on GPQA Diamond, 32.8 on LiveCodeBench, 69.4 on MMMU, 88.8 on ChartQA, 94.4 on DocVQA and 70.7 on MathVista.^[2] A benchmark is a standard test set, so these scores are useful for comparing models on the same task, not for predicting every real use.

Image grounding. Meta described Scout as best-in-class for image grounding, meaning the task of linking words in a request to the matching parts of an image.^[1]

Languages. Scout officially supports 12 languages: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese.^[2]

Licence and access. Scout was released under the Llama 4 Community License Agreement and can be downloaded from llama.com and Hugging Face.^[2][1] It was also offered in Meta's Llama API limited preview, announced at LlamaCon on 29 April 2025.^[3]

Meta AI. Meta AI on WhatsApp, Messenger, Instagram Direct and meta.ai was rebuilt on Llama 4.^[1]

Siblings. Scout was released alongside Llama 4 Maverick. Llama 4 Behemoth was previewed but not released.^[1]

Its place in development

Scout's main departure from earlier Llama models is its context length. Going from 128K to 10 million tokens lets a single request contain far more text, images or video than before.

Scout and Maverick together show Meta's move to mixture-of-experts designs. The larger Maverick model shares the same family and was released the same day.^[1]