Llama 4 Scout

Meta's smaller Llama 4 model: a natively multimodal Mixture-of-Experts model with 109B total / 17B active parameters (16 experts) trained on ~40T tokens. Its headline feature is a 10M-token native context window — the longest shipped in an open-weight model at release — enabled by an iRoPE attention design (interleaved local/global attention). Accepts text and image input; multilingual (12 supported languages); August 2024 knowledge cutoff. Fits on a single H100 via int4 quantization. Weights gated behind the Llama 4 Community License on Hugging Face and llama.com; also powers Meta AI on WhatsApp/Instagram/Messenger. Note: this is Scout — the 17B-active model — not Maverick (400B total, 1M context) or the unreleased Behemoth teacher.

Provider-reported (Meta, Apr 2025): best multimodal model in its class vs Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1. Independent (RedHatAI evals on meta-llama/Llama-4-Scout-17B-16E-Instruct, FP8): MMLU 80.54, IFEval 86.90, MMMU 53.44, OpenLLM v1 avg 77.48, RULER avg 80.44. Caution: Meta's launch LMArena score was achieved with an unreleased 'experimental chat version' optimized for conversationality rather than the public checkpoint — LMArena later changed its policies over this incident.