Llama 4 Scout: Meta's 10M-Token Open-Weight Model

What is Llama 4 Scout?

Llama 4 Scout is an open-weight multimodal model from Meta, announced 5 April 2025, with a 10 million token context window - the largest of any widely released model we could find. It uses a mixture-of-experts architecture with 17 billion active parameters and 109 billion total across 16 experts. Meta's announcement does not name a specific licence, so check the terms on llama.com or Hugging Face before commercial use - an earlier version of this page asserted "free commercial use" without that being stated anywhere.

The 10 Million Token Context Window

A 10 million token context window is a step change: it is enough to hold entire codebases, shelves of books, or years of documents in a single prompt without retrieval tricks. For use cases like whole-repository code understanding, large-scale document review, and long-horizon agents that must remember everything they have seen, Scout removes constraints that force other models into complex chunking and retrieval pipelines. This alone makes Llama 4 Scout uniquely capable for certain workloads.

Architecture and Efficiency

Llama 4 Scout uses an efficient mixture-of-experts architecture that activates only a fraction of its parameters per token, keeping inference cost and latency manageable even with its enormous context capacity. It is natively multimodal, handling text and images, and is designed to be practical to deploy - Meta and the community provide quantized variants that run on a range of hardware, from multi-GPU servers down to well-specified workstations for smaller configurations.

Licensing: Check Before You Deploy

Correction (12 August 2026): this section previously stated the model was "free to download and use commercially under Meta's community license". Meta's own announcement names no specific licence. It is downloadable from llama.com and Hugging Face, but you should read the actual terms on those pages before commercial use rather than relying on this guide's earlier, unsourced claim.

Ideal Use Cases

Based on its verified 10 million token context window, Scout is a reasonable choice for workloads that need to hold very large inputs at once - whole codebases, large document sets, long transcripts - without a retrieval pipeline. It now appears on the LMArena leaderboard too (1322 +/-5, rank 237 as of 23 August 2026), though near the bottom - a ranking that mostly reflects its older, non-reasoning design rather than what the context window makes possible.

Limitations

Fully exploiting the 10M token window requires substantial memory and careful engineering; most deployments run at smaller effective context to control cost. On head-to-head reasoning and coding benchmarks Scout is strong but not always the open-weight leader, and its multimodal breadth trails the top closed models. Self-hosting demands MLOps capability, and as with all models, outputs should be verified for factual and safety-critical use.

Llama 4 Scout vs Competitors

On context window alone, no other model on this site is close: 10 million tokens against Claude Opus 5 and Gemini 3.1 Pro at 1 million, GPT-5.6 at 1.05 million, and Mistral Large 3 at 256k. It was absent from the LMArena text leaderboard when this guide was first written; as of 23 August 2026 it is listed (llama-4-scout-17b-16e-instruct) at rank 237 with an Arena score of 1322 +/-5 - far below every closed model covered here, so weigh the unmatched context window against a much weaker human-preference ranking.

Official Resources