Llama 4 Scout

Scout is a sparse MoE transformer: 109B total parameters across 16 routed experts plus a shared expert, ~17B active per token. It uses the iRoPE attention scheme — chunked RoPE local attention (8K chunks) in three of four layers and a full-context NoPE (no positional embedding) layer every fourth, which is what makes the 10M window mechanically possible. Scout adds QK-normalization (RMS norm of query/key states, no learnable params) and temperature-scaled softmax in NoPE layers to preserve attention over very long sequences. Vision is early-fused. Meta discloses total/active params, expert count, the iRoPE design, and a training-token scale of up to ~40T; exact layer counts, compute, and full data recipe are not published. The released checkpoint is Instruct-tuned and is the only Llama 4 variant designed to fit a single server-grade GPU.

Scout is a competent generalist with one standout trick — context length. Text reasoning (cap_reasoning 5.5) is roughly on par with Llama 3.1 70B: MMLU 79.6, MMLU-Pro 74.3, GPQA Diamond 57.2 — solid for its tier but not frontier. Math (5.0) and coding (5.5) are mid-pack open-weights; HumanEval ~82, LiveCodeBench ~33. Multilingual (7.5) is a genuine strength across 200 languages. Vision and document OCR (6.5 / 7.0) are strong for the size, with ChartQA 85.3, DocVQA 91.6, and MMMU 69.4. The long-context score (4.5) is the honest tension at the heart of this model: needle-in-haystack retrieval over the 10M window looks near-perfect, but independent comprehension benchmarks (Fiction.LiveBench, RULER-style) show meaningful quality degradation long before the ceiling — the 10M number is a retrieval capacity, not a reasoning-over-10M guarantee. Function-calling and instruction-following (6.0 each) are reliable for single steps, weaker on long agent chains. No reasoning mode, no real-time data (0.0).

Median output speed is ~106 tokens/sec across providers, with time-to-first-token around 0.56s on DeepInfra and 0.72s on Google Vertex. Because only ~17B params are active, it generates at small-model speed despite the 109B pool. Groq is the throughput leader at ~449 tokens/sec, giving a snappy sub-second interactive feel. Latency tier is fast; the practical caveat is that filling the 10M context dramatically raises prefill latency and cost, so the long-context superpower is best used selectively.

Open weights mean no single Meta price; the figures above are the May 2026 inference market. Scout is the cheapest serious open-weights multimodal model on most providers.

Open weights under the Llama 4 Community License. Download from Hugging Face (meta-llama/Llama-4-Scout-17B-16E-Instruct). The headline deployment property: it fits a single server-grade GPU. At INT4 the full 109B parameter set needs roughly 55–60GB, so one H100 80GB serves it comfortably; aggressive GGUF quants (Q4) run on 24–48GB consumer cards for hobbyist/edge use, though storing all 109B params is what sets the floor. Managed availability spans AWS Bedrock, Google Vertex AI, Azure AI Foundry, OCI, and IBM watsonx. Inference providers include Together, Fireworks, Groq, DeepInfra, OpenRouter, and Novita. Self-host economics are the strongest selling point — a single rented H100 at $2–3/hour serves millions of tokens/day. The Llama 4 Community License permits commercial use but requires a separate Meta license above 700M MAU and forbids training non-Llama models on outputs.

Identical posture to Maverick: the weights carry no built-in moderation, and Meta offers Llama Guard 4 (12B multimodal) plus Prompt Guard 2 (22M/86M) as optional pre/post filters for the Llama 4 line. "Trains on inputs" is not applicable when self-hosted; managed-provider terms vary; Meta's own terms do not train on your data. No model-level compliance certifications — those attach to your host or infrastructure. Refusal calibration is moderate and tunable, which is the intended advantage for regulated and sovereign deployments. Governance under Meta's Frontier AI Framework.

Native support across Hugging Face Transformers, vLLM, llama.cpp, Ollama, SGLang, MLX, plus LangChain and LlamaIndex. Available on Bedrock, Vertex AI, Azure AI Foundry, OCI, and IBM watsonx, and on Together, Fireworks, Groq, DeepInfra, OpenRouter, and Novita. Used inside Meta's consumer AI surfaces. Popularity is mainstream — the go-to open-weights pick when single-GPU deploy or very large context is the requirement.