Open-weight 17B-active mixture-of-experts model with a 10M-token context window and image input for large-document and multi-file analysis.
Modality:
Text
Image
model ID
meta-llama/llama-4-scout
Output Speed *
56.18
tok/s
Intelligence Index *
8.1
/ 100
Context Window *
327680
tokens
Input price
0.6
Anytoken
Output price
1.8
Anytoken
Llama 4 Scout: 10M-Token Context in a 17B-Active Open-Weight Model
Llama 4 Scout is Meta's smaller Llama 4 model. It is a mixture-of-experts system with 17B active parameters, 16 experts and 109B total parameters. Released with open weights in April 2025, it accepts text and image input and lists a 10M-token context window, the largest in the Llama 4 family. It sits below Llama 4 Maverick in reasoning and coding. It suits workloads that must read large codebases, document collections or long histories in a single request.
Access Llama 4 Scout via the AnyAPI.ai API
Performance
Long-Context Reach and Document Vision at 17B Active Parameters
Scout's differentiator is input scale, not peak reasoning. Meta reports a 10M-token window, MMMU 69.4 and DocVQA 94.4 for the instruction-tuned model. It also reports MMLU Pro 74.3 and LiveCodeBench 32.8. These are Meta-reported figures. Our interpretation is that document and chart understanding is strong for the model's active size, while coding and graduate-level reasoning trail larger models. In production, use Scout to ingest very large inputs, and route hard coding or reasoning steps to a stronger model.
Technical Specifications
What the model supports
Scout takes text and image input and returns text and code. Meta lists a 10M-token context window. The model was pretrained and post-trained at 256K, and longer lengths rely on length generalization, so test quality at your real prompt size. Hosted endpoints may cap context below 10M. The context window is not the output limit. No single official maximum-output figure applies across hosts, so check the limit of the endpoint you call.
Verified Specifications —
Llama 4 Scout
*
Input modalities
Text
Image
output modalities
Text
Context window
327680
tokens
Maximum output tokens
16384
Reasoning
No
Knowledge cutoff
April 2025
Pricing (standard)
0.6
AnyTokens in
/
1.8
AnyTokens out
Comparison
Llama 4 Scout or Llama 4 Maverick: What Actually Changes?
Scout and Maverick are Meta's two Llama 4 models, and both are multimodal mixture-of-experts models with 17B active parameters. They differ in scale and context. Scout has 16 experts, 109B total parameters and a 10M-token window. Maverick has 128 experts, about 400B total parameters and a 1M-token window. Meta's reported scores favor Maverick: MMLU Pro 80.5 against 74.3, GPQA Diamond 69.8 against 57.2, and LiveCodeBench 43.4 against 32.8. The decision is context reach against reasoning and coding quality.
Dimension
Llama 4 Scout
Llama 4 Maverick
Context window *
327680
tokens
1048576
tokens
Output speed *
56.18
tok/s
34.36
tok/s
Intelligence Index *
8.1
10
Input pricing
0.6
AnyToken
0.9
AnyToken
Output pricing
1.8
AnyToken
3.6
AnyToken
Knowledge cutoff *
April 2025
April 2025
Choose Llama 4 Scout when a request must hold the largest possible input, such as a full codebase or a large document set. Also choose it for a lighter, lower-cost Llama 4 tier on document and image understanding, where Meta's DocVQA score matches Maverick's. Choose Llama 4 Maverick when output quality matters more than input size: coding, harder reasoning, and stronger image reasoning (MMMU 73.4 against 69.4). Maverick is also the safer choice if 1M tokens already covers your prompts.
Limitations & Trade-offs
1
Weaker coding and reasoning. Meta reports LiveCodeBench 32.8 and GPQA Diamond 57.2 for Scout, below Llama 4 Maverick's 43.4 and 69.8. Scout has no reasoning mode. Code generation, multi-step debugging and expert-level question answering are therefore better handled by Maverick or a dedicated reasoning model. Scout suits them only when the input size, not the task difficulty, is the constraint.
2
The 10M window is not proven quality at 10M. Meta states Scout was pretrained and post-trained at 256K context, with length generalization beyond that. Hosted endpoints may also cap context well below 10M. Very large prompts add latency and cost. For retrieval-style questions over huge corpora, a RAG pipeline with a smaller prompt may be more reliable. Validate accuracy on your own long inputs before depending on the maximum window.
3
Text-only output and limited vision scope. Scout accepts text and images but generates only text and code. There is no audio, video or image generation. Meta tested image understanding with up to five input images and lists English only for image tasks. Image-heavy workflows with many pages per request or non-English visual content need a separate check or a different model.
4
Narrow language coverage and a fixed knowledge cutoff. Meta lists 12 supported text languages. Other languages may work but fall outside what Meta supports. The knowledge cutoff is August 2024, so Scout lacks later events and library releases unless you supply them in the prompt. Applications needing broad language coverage or current knowledge should add retrieval or use a model with a later cutoff.
Best-Fit Workloads
Where this model earns its place
01
Whole-Codebase and Repository Analysis
Scout's 10M-token window is the largest in the Llama 4 family, so a large repository, its documentation and its issue history can fit in one request. That suits code search, architecture summaries and cross-file dependency questions. Its LiveCodeBench score is modest, so use it to read and explain code, and send complex code generation to a stronger coding model. Confirm your endpoint's context limit first.02
Large Document Set Review
Legal bundles, research collections and long reports can be analyzed together instead of chunked. This lets the model compare clauses or findings across documents in one pass. Meta's MMLU Pro score of 74.3 indicates solid general knowledge. Test recall on your real prompt sizes, since the model was trained at 256K. For narrow lookups over a huge corpus, pair it with retrieval.03
Document, Chart and Screenshot Understanding
Scout accepts images alongside text. Meta reports DocVQA 94.4, ChartQA 88.8 and MMMU 69.4, which supports invoice extraction, chart reading and UI screenshot analysis. Meta lists English only for image understanding and tested up to five images, so avoid multilingual or many-page scan pipelines without your own evaluation. Output is text only.04
Multilingual Assistants and Triage Tier
With 17B active parameters and support for 12 text languages, Scout works as a first-pass tier for support triage, classification and summarization across those languages. Escalate difficult coding or reasoning requests to Maverick or another stronger model. The August 2024 knowledge cutoff means current facts should come from retrieved context.Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04