Llama 4 Scout

Llama 4 Scout, the Llama-4-Scout-17B-16E-Instruct checkpoint with 17 billion active and 109 billion total parameters, was released by Meta on 5 April 2025. It is the newest Llama-branded model in what we read: two reference pages say Meta moved its main assistant to the closed Muse Spark in April 2026, and neither lists a later Llama. The question here is what developers say they use Scout for beyond its headline.

We read 14 sources on 9 October 2026, dated 5 April 2025 to 13 July 2026. Four are official Meta, Hugging Face and licence pages. Two are an evaluation site and a technology publication. One is a news report about the sibling Maverick model. Five are independent engineering guides and reference pages, and two are developer-forum sources. There is no buyer tier, and the editorial tier is only two pieces, so every count below is small.

Consensus

Across the sources we read, Scout is described as a reader of documents and images, as cheap to call, and as runnable on one large GPU or a high-memory desktop once quantised. The sources dispute the part Meta leads with: the 10-million-token context window. This is our reading of the sample on 9 October 2026, not a measurement of the model.

Confidence is mixed. The facts about size, date and context length come from Meta's own pages, which are the maker's claims. Independent evidence is thin: two evaluation or editorial sources, one dated April 2025.

Recurring strengths

Reading long documents and whole codebases. Three of the 14 sources name it: Meta's launch post (the maker's claim, listing multi-document summarisation and reasoning over large codebases), and two engineering guides dated 8 April and 13 July 2026, one naming document processing and the other whole-codebase analysis and combining many sources in one pass. Neither evaluation source names it as a use.

Reading charts, documents and images. Three sources name it: Meta's model card lists multimodal understanding, one guide names chart interpretation and visual question answering, and the other names text, image and video workflows. The model card adds that image understanding is limited to English.

Cheap to call as a hosted model. Both evaluation or editorial sources point to price. The evaluation site's page, read on 9 October 2026, lists $0.19 per million input tokens and $0.68 per million output tokens, and ranks Scout at 8 on its Intelligence Index, above the 7 it gives as the median of comparable models. The Decoder reported in April 2025 that Scout costs roughly a tenth of GPT-4o. A price is not a quality score.

Runs on one large GPU or a high-memory desktop once quantised. Four sources say so in different terms: the model card (a single H100 with INT4 weights), one guide (48 to 64 GB of VRAM, or a Mac with 64 GB or more, at roughly 12 to 18 tokens a second on those Macs), the other guide (24 GB of VRAM with aggressive quantisation, around 20 tokens a second), and one developer comment reporting almost 20 tokens a second on a Strix Halo machine. The speeds are the authors' own reports.

For scale, the Hugging Face page for Scout showed 170,153 downloads in the last month and 1,368 likes on 9 October 2026. Downloads say how many people fetched the weights, not how many use them.

Recurring complaints

The long-context window does not hold up. Four of the 14 sources raise it. The Decoder reported in April 2025 that on the Fiction.Live benchmark Scout reached 15.6 percent accuracy on narratives of 120,000 tokens. One guide repeats a 15.6 percent figure at 128,000 tokens and puts reliable use near 32,000 tokens. A Verdent guide says the single-H100 claim does not show the full window can be served at production latency, and notes pretraining used 256,000 tokens. In the launch thread on Hacker News, several commenters questioned recall across the whole window. No source counts how many developers hit this, so how common it is stays unknown.

The memory bill. Three sources raise it: a launch-thread commenter putting the minimum near 54.5 GB of VRAM, one guide giving 48 to 64 GB for a quantised setup, and the other guide saying the full 10-million-token window needs eight H100 cards and a budget above $60,000.

Weak for coding. Two sources say so: one guide, which gives Scout 32.8 percent on LiveCodeBench against 33.3 percent for Llama 3.3 70B, and one developer comment saying it fixes bugs by commenting code out and adding TODOs. That comment is one developer's opinion.

Where reviewers split

Whether the long window is a strength. Meta's launch post and one July 2026 guide present it as the reason to pick Scout. The Decoder and the other guide report it failing well before 10 million tokens. We read this as a split between what the window allows and what the benchmark found, and neither side measured the other's case.

General ranking against coding. The evaluation site's index places Scout above its peer median, while one guide calls it unsuitable for coding on one benchmark. These are different tests of different things and both can be true.

A different model, cited once. TechCrunch reported on 11 April 2025 that the public Maverick ranked 32nd on LM Arena, below several older rivals, after Meta had used a tuned experimental version in its launch chart. That is Maverick, not Scout, and is not counted anywhere above.

Individual opinions. A handful of developer comments describe a concrete use, such as local refactoring on a Mac Studio or long-context tests on a private dataset. Each is one developer's account.

Who it suits

Developers with a stack of documents, charts or screenshots to read in bulk, who can either pay a hosted provider a low per-token price or run a quantised copy on a machine with 48 GB or more of memory. Developers who want a long input but will check the answer at lengths past about 32,000 tokens, since the sources we read disagree on how far recall holds.

Who should pass

Anyone who needs dependable recall across hundreds of thousands of tokens, since the sources we read do not support that. Anyone whose main job is writing code, since the two sources that mention coding both fault it. Anyone who wants a model that is still being developed, since the Llama line stopped at this release in what we read. The licence is Meta's Llama 4 Community License Agreement, effective 5 April 2025. This review does not say whether it allows your use, and the licence file is the place to read it.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it. Get the weekly verdict

Sources

• Meta, Llama 4 launch post, official, 5 April 2025
• Meta, Llama 4 model card, official, read 9 October 2026
• Hugging Face, Llama-4-Scout-17B-16E-Instruct, official, read 9 October 2026
• Llama 4 Community License Agreement, official, effective 5 April 2025
• Artificial Analysis, Llama 4 Scout, evaluation, read 9 October 2026
• The Decoder, Llama 4 on standard tests and long context, editorial, 7 April 2025
• TechCrunch, vanilla Maverick ranks below rivals, news, 11 April 2025
• Zen van Riel, Llama 4 Scout practical guide, independent guide, 8 April 2026
• DEV Community, Llama 4 Scout locally, independent guide, 13 July 2026
• Verdent, Llama 4 Scout guide, independent guide, undated, read 9 October 2026
• Wikipedia, Llama (language model), reference, read 9 October 2026
• Mungomash, Llama versions, reference, updated 29 September 2026
• Hacker News, The Llama 4 herd, forum, 5 April 2025, partial view of an ordinary page
• Hacker News search, comments on Llama 4 Scout, forum, one API call, read 9 October 2026

Fourteen sources dated 5 April 2025 to 13 July 2026 were read on 9 October 2026 and synthesised with AI, and no model was run and nothing was tested by us.