The context window is 10 million tokens. That number is technically accurate. It is also, for most workplace deployments, irrelevant - and understanding why tells you nearly everything about how to use Llama 4 Scout well.
Scout (17B active / 109B total parameters, 10M context) launched April 5, 2025 , and the headline grabbed attention immediately. A 10M-token context means you could feed the model roughly 75 full-length novels in a single prompt. For teams managing large codebases, sprawling knowledge bases, or months of meeting transcripts, that sounds like it dissolves whole categories of retrieval infrastructure.
The reality is more specific - and more useful once you know the shape of it.
What the 10M context actually costs in GPU memory
The KV cache is the part of the equation that vendor announcements skip. Scout supports up to 10 million tokens of native context. That number is not primarily a compute problem - it is a memory problem. Scaling linearly, a single Scout inference at 10M tokens requires approximately 500GB of KV cache.
To put that in GPU terms: Scout at INT4 (55GB) fits on a single H100 80GB with 25GB left for KV cache, which is tight but functional at 4K-32K context. Push toward 128K tokens and the KV cache at 128K context runs to roughly 24GB , which consumes essentially all the headroom on that single card. At 1M tokens, most hosted providers cap the practical context length well below Scout's theoretical 10 million for exactly this reason.
The arithmetic matters more than the spec sheet:
| Context length | KV cache (Scout, FP16) | Viable on single H100? |
|---|---|---|
| 32K tokens | ~1.6 GB | Yes (INT4 weights) |
| 128K tokens | ~6.4 GB | Yes (INT4 weights) |
| 1M tokens | ~50 GB | No - needs H200 or multi-GPU |
| 10M tokens | ~500 GB | No - far exceeds any single card |
Sources: Clusterbid VRAM analysis, Hardwarepedia Scout requirements
Where the context window is genuinely useful - and where it breaks
Scout's context ceiling earns its headlines in one specific scenario: retrieval without a retrieval pipeline. The massive capacity enables native RAG-less workflows where entire technical libraries, thousands of high-resolution images, or massive datasets can be loaded directly into the KV cache for immediate reasoning.
That is real. Feeding a 400-page PDF or an entire mid-size codebase into a single prompt is a workflow change with actual consequences: no chunking logic, no embedding freshness problem, no retrieval misses from poor vector similarity.
The breakage point is harder reasoning over long spans. Effective recall starts collapsing past approximately 1M tokens. Four-bit quantization - required for consumer-GPU deployment - adds 15-20% perplexity past 5M tokens. Third-party benchmarks make the gap concrete: on Fiction.Livebench, which measures whether a model can answer questions requiring understanding of relationships between distant parts of a 128K context, Scout scores 15.6% against Gemini 2.5 Pro's 90.6%.
The usable split for workplace tasks looks roughly like this:
- Legal or scientific literature review - loading a full corpus for citation lookup: Scout handles this well at 128K-500K tokens
- Codebase-wide code review or security audit - single-pass over a monorepo: practical up to ~200K tokens, degrades on multi-file reasoning above that
- Multi-hop questions over months of Slack exports - requires connecting many scattered facts: break this into targeted retrieval instead; Scout's long context will not reliably stitch together distant signals
- Meeting transcript summarization - 50-100K tokens at a time: genuinely good, no retrieval plumbing needed
Beagle in action#engineering, 11:02amThe case for Scout over a closed API - and what it actually costs
At the context lengths where Scout works well (32K-128K), the economics are straightforward. Scout's API price via third-party providers runs around $0.18 per million input tokens and $0.66 per million output tokens. That is already cheaper than GPT-4o-class APIs for the same task. Self-hosting on a leased single H100 changes the math further: a Scout INT4 deployment on one H100 runs roughly $1,440 per month at 24/7 uptime , which breaks even against API pricing at somewhere around 2 billion tokens per month - achievable for a team running continuous document analysis or automated code review pipelines.
The Llama 4 Community License permits commercial use for everyone except hyperscalers above 700M MAU , so most enterprise teams can self-host without a legal review conversation. Day-zero support landed across vLLM, Ollama, llama.cpp, SGLang, Transformers, and most agentic frameworks.
The honest non-obvious point: Scout's 10M context window is not what makes it interesting for teams. What makes it interesting is that MoE architecture means all 109B parameters must sit in memory, but each token only activates 17B of them - memory capacity requirements are those of a 109B model, while speed is that of a 17B model. At 128K context on a single H100, you get Llama 3.3 70B-class intelligence at closer to Llama 3.1 8B-class inference speed. That is the actual trade teams should be making the decision around.
Llama 4 Scout 10M context: common questions
What is Llama 4 Scout's actual usable context length?
The 10M ceiling is technically real for exact retrieval tasks. In practice, reasoning quality degrades meaningfully past 500K tokens, and hardware constraints cap most single-GPU deployments at 128K tokens before running out of KV cache headroom. For workplace workloads, treat 128K as the reliable ceiling.
Can Llama 4 Scout run on a single GPU?
Yes, with INT4 quantization. At Q4_K_M, the weights fit in roughly 55-68GB of VRAM, making a single H100 80GB viable for contexts up to about 32K-128K tokens. Beyond that, KV cache growth pushes you to a second GPU or an H200.
Does Llama 4 Scout replace a vector search / RAG pipeline?
For document sets under roughly 100K tokens, yes - you can load them directly into context and skip chunking and embedding entirely. For larger corpora or questions that require connecting distant facts, a retrieval layer still performs better on multi-hop tasks.
How does Llama 4 Scout compare to Maverick for team use?
Scout is the better fit for most workplace deployments: it has the larger context window (10M vs. 1M), fits on a single H100 at INT4, and runs faster at inference. Maverick's 128 experts give it a quality edge on hard reasoning but require a full DGX host (8× H100) to run, which is a significant infrastructure commitment.
What is the Llama 4 Scout license for commercial use?
The Llama 4 Community License allows commercial use for any company with fewer than 700 million monthly active users. Most enterprise teams qualify without restriction. You can download the weights from Meta's official model page or via Hugging Face.