Llama 4 Scout

Llama 4 Scout is built on a mixture-of-experts (MoE) architecture that activates 17 billion parameters from a pool of 109B total, using 16 experts per forward pass. This design allows it to deliver strong performance while remaining inference-efficient, making it suitable for both research and commercial deployment. The model supports native multimodal input—text and images—with early fusion enabling seamless modality integration. With an exceptional 10 million token context window, it handles long documents, multi-turn conversations, and complex visual reasoning tasks that smaller models cannot easily manage. Scout is instruction-tuned for assistant-style interaction, excelling at multilingual chat, image captioning, and visual understanding across 12 languages.

Trained on approximately 40 trillion tokens with a knowledge cutoff of August 2024, Llama 4 Scout benefits from Meta's extensive pre-training scale. The model incorporates open-sourced safety tools—Prompt Guard and Llama Guard—along with GOAT (Generative Offensive Agent Testing), an automated red teaming framework that improves safety alignment. Benchmark results show Scout outperforming comparable models like Gemma 3 27B and Gemini 2.0 Flash Lite on coding, reasoning, and image tasks. Available under the Llama 4 Community License with open weights, it offers flexibility for local deployment or commercial use, with routing through OpenRouter for accessibility.

OpenRoutermeta-llama/llama-4-scoutllama

Open provider page Provider docs

Quick Info

Provider: OpenRouter
Model key: meta-llama/llama-4-scout
Release date: Apr 5, 2025
Last updated: Apr 5, 2025
Knowledge cutoff: 2024-08-31
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $0.30

Limits

Output tokens: 16,384 tokens
Context window: 1,310,720 tokens

Transparent token rates

Compare Llama 4 Scout pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Llama 4 Scout

OpenRouter

CoverageAnalysis

A April 6, 2025 technical analysis details Llama 4 Scout's mixture-of-experts implementation, which is Meta's first MoE architecture. Scout has 17B active parameters and 16 experts per forward pass, drawn from a total of 109B parameters. The model is pre-trained and post-trained with a 256K context length and features early-fusion multimodal integration across text, image, and video modalities. This design enables length generalization and efficient parameter activation per token. The comparative analysis positions Scout against other leading multimodal LLMs on benchmarks including MMLU, GSM8K, HumanEval, and MMMU. Scout's MoE approach selectively activates only a subset of expert parameters per token via a routing mechanism, a departure from dense transformer designs used in prior Llama generations. Community authors reference Meta's official blog at ai.meta.com/blog/llama-4-multimodal-intelligence as the source for architectural details.

OpenRouter

Coverage

IBM announced on April 7, 2025 that Meta's Llama 4 Scout is now available on the watsonx.ai platform alongside Llama 4 Maverick. The announcement notes these are Meta's first mixture-of-experts (MoE) models, providing frontier multimodal performance, high speeds, low cost, and industry-leading context length. Both models support text-in/text-out and image-in/text-out use cases, and their addition brings IBM's library of supported Meta models to 13 on watsonx.ai. The IBM release frames Llama 4 Scout as part of Meta's new architectural direction, integrating text, image, and video modalities earlier in training than conventional models. Available under the Llama 4 Community License on watsonx.ai, Scout enables enterprise customers to access the model's multimodal capabilities and long context length for deployment. The announcement underscores IBM's open, multi-model strategy by adding frontier open-weight models alongside existing Meta offerings.