Llama 4 Scout: What the 10M Token Context Means in Practice

Llama 4 Scout is remembered for one number: a 10 million token context window, announced by Meta on 5 April 2025 as the longest in the industry. That number is real in the sense that the architecture supports it. It is misleading in the sense that almost nobody has used it, and the model was never trained on sequences anywhere near that long.

Scout is still a useful model. It is a compact multimodal MoE that fits on one GPU when quantised and reads documents as well as its larger sibling. This guide separates what the context claim means from what Scout is actually good for.

The specs

Scout has 109B total parameters and 17B active per token, across 16 experts. Like Llama 4 Maverick, it is natively multimodal: text and images go in, text comes out. Meta lists twelve supported languages, around 40 trillion training tokens, and an August 2024 data cutoff.

Meta released it in BF16 and says it fits on a single H100 with on-the-fly int4 quantisation. That is the main practical difference from Maverick, which needs a full eight-GPU host even in FP8. Scout also runs on large-memory Macs: one early test on an M3 Ultra reported about 53 tokens per second at 3-bit with 47 GB of memory in use, falling to about 12 tokens per second at fp16 with 216 GB. It does not fit on consumer graphics cards, even quantised.

Where the 10M figure comes from

Meta says Scout was pretrained and post-trained with a 256K context length. The jump to 10M comes from the architecture: interleaved attention layers that use no positional embeddings, which Meta calls iRoPE, intended to let the model generalise to lengths far beyond what it saw in training.

Generalising and performing well are not the same thing. Every token beyond 256K is a length the model never saw during training. Retrieval of a single fact may hold up; reasoning across many scattered facts usually degrades well before the advertised limit. Long context degradation explains the pattern, and Scout is a clear example of it.

There is also a hardware ceiling. Meta's own example notebook noted that on eight H100s in BF16 you can reach about 1.4M tokens. The KV cache for 10M tokens simply does not fit on ordinary serving hardware, which is why the full window exists mainly on paper.

What hosts actually serve

At launch, providers capped Scout at 128K (Groq and Fireworks) or 328K (Together). Current hosted listings are larger in some places, but none we have seen offers anything close to 10M. The figure that matters is the one on the endpoint you call, and it varies by provider and sometimes by plan.

The honest summary: treat Scout as a model with a long and somewhat elastic context that is solid up to its training length and less predictable beyond it. If a 1M-class window is the requirement, the 1M-context comparison covers which windows hold up under testing.

What it is good at

Meta's instruct benchmarks show where Scout earns its place:

• DocVQA — 94.4, identical to Maverick
• ChartQA — 88.8, against 90.0 for Maverick
• MMMU — 69.4
• MathVista — 70.7
• MMLU Pro — 74.3, against 68.9 for Llama 3.3 70B
• GPQA Diamond — 57.2, against 50.5 for Llama 3.3 70B
• LiveCodeBench — 32.8, against 33.3 for Llama 3.3 70B

Reading scanned documents, forms and charts is the standout: on DocVQA Scout matches a model with nearly four times its total parameters. Against Llama 3.3 70B it is ahead on knowledge and reasoning and level on code, while also accepting images. Meta notes it tested image understanding with up to five images per request, so treat larger image batches as untested territory.

Code is the weak spot relative to Maverick, which scores 43.4 on the same LiveCodeBench window. If coding quality drives your choice, Scout is not the Llama 4 to pick.

Cost and licence

OpenRouter lists Scout at $0.10 per million input tokens and $0.30 per million output at the time of writing. That is roughly what it lists for Llama 3.3 70B, so through an API the comparison is about capability, not price. Note that long context multiplies input cost: a single 300K-token prompt costs the same as three hundred 1,000-token prompts.

The licence is the Llama 4 Community License Agreement, with the standard terms (the 700 million monthly active user threshold, "Built with Llama" attribution and the naming rule for derived models). Meta's FAQ also states that rights to the Llama 4 multimodal models are not granted to EU-domiciled individuals or companies, though end users of products built on them are not affected. Scout is multimodal, so EU teams should check this before self-hosting.

How to test the long context yourself

1. Confirm the provider's actual context limit and maximum output first. Do not trust the model card for this.
2. Build test documents at 32K, 128K, 256K and the largest size the host allows, from your own material.
3. Ask questions that need one fact, then questions that need three facts from different parts of the document.
4. Plot accuracy against length for each question type. Expect single-fact retrieval to hold longer than multi-fact reasoning.
5. Compare against chunking plus retrieval on the same questions. If retrieval matches the long-context result, it is cheaper and you should use it.

Choose Scout when you want an inexpensive multimodal model you can self-host on one GPU, mainly for document and chart work. Choose Maverick for stronger reasoning and code. Choose neither if a dependable multi-million-token window is the core requirement, because no host currently delivers one with Scout.