Meta just dropped a 17 billion parameter model with a 10 million token context window and called it edge-optimized. That’s the pitch. The reality is messier.<\/p>\n
Llama 4 Scout launched in April 2025 as Meta’s answer to the long-context arms race. While Gemini and Claude topped out at roughly 1 million tokens, Meta went open-source with 10x that capacity. The model uses a mixture-of-experts architecture with 109 billion total parameters but only activates 17 billion per inference pass, which explains how it runs faster than you’d expect from the raw parameter count. It handles text and vision. It’s free to download. And it requires hardware that most developers don’t own.<\/p>\n
This guide exists because Llama 4 Scout matters for a specific reason. It’s the first credible open-source challenge to proprietary long-context models. If you need to process entire codebases, analyze hundreds of research papers simultaneously, or run AI on sensitive data that can’t touch the cloud, this is the model that makes it possible without a seven-figure API bill. But the “edge” label is marketing fiction. True 10 million token processing requires eight Nvidia H100 GPUs, which costs more than most cars. The model is remarkable. The positioning is dishonest.<\/p>\n
You should care about Scout if you’re building local AI infrastructure, if you handle data that regulatory frameworks won’t let you upload to OpenAI, or if you’re tired of per-token pricing that scales linearly with your success. You should ignore it if you need state-of-the-art reasoning, if you’re doing agentic coding workflows, or if you don’t have at least one serious GPU.<\/p>\n
The model uses mixture-of-experts architecture, which means it has 109 billion parameters total but only routes each token through 17 billion of them. This is how it achieves faster inference than dense models of similar capability. The 10 million token context window is real, verified by Meta’s official announcement<\/a> and confirmed in testing by multiple cloud providers. For reference, 10 million tokens equals roughly 7,500 pages of text or 2.5 million lines of code.<\/p>\n Meta released Scout alongside larger Llama 4 variants in April 2025. The model supports both text and vision inputs, making it genuinely multimodal rather than text-only with image understanding bolted on. You can feed it screenshots, diagrams, charts, and PDFs with visual elements. The vision capabilities scored competitively on MMMU and ChartQA benchmarks, though Meta hasn’t published detailed vision-specific performance data.<\/p>\n The catch is infrastructure. DeepLearning.AI reported<\/a> that running the model at scale requires eight H100 GPUs for contexts beyond 1.4 million tokens. That’s $60,000 in hardware at minimum. For contexts under 500,000 tokens, you can get away with a single A100 or even a consumer RTX 4090 with quantization. But the headline 10M capability needs data center resources.<\/p>\n Meta distributed the model under their standard Llama Community License, which allows commercial use with some restrictions for companies over 700 million monthly active users. The weights are available on AWS<\/a>, Hugging Face, and through Cloudflare Workers AI<\/a>. No official API exists. You download the weights and run them yourself or pay a cloud provider to host them for you.<\/p>\nSpecs at a glance<\/h2>\n