Llama 4 Scout is a general purpose model with 17B active parameters, 16 experts, and 109B total parameters. Features an industry-leading 10M token context length, enabling multi-document summarization, parsing extensive user activity, and reasoning over vast codebases.
Llama 4 Scout is Meta's open-weight multimodal MoE model for extreme-context workloads. It was released with the Llama 4 family on April 5, 2025, activates 17B parameters per forward pass across 16 experts, and is listed by Meta with 109B total parameters, native text-and-image input, text-and-code output, and a 10M token context window.
Use Scout when the routing constraint is context span more than peak reasoning quality: full-book translation, repository-scale inspection, long evidence packets, and multimodal document review. Compared with Llama 4 Maverick, Scout trades fewer experts and lower benchmark scores for a context window ten times larger. Through TheRouter it is a managed OpenAI-compatible route rather than Meta's raw checkpoint workflow.
Meta publishes Llama 4 Scout as open weights under the Llama 4 Community License, with BF16 weights and support for on-the-fly int4 quantization in Meta's release notes.
TheRouter serves Scout as a managed OpenAI-compatible model route with catalog pricing, context, modalities, and supported parameters coming from standard-models.yaml.
| Type | Rate |
|---|---|
| Input | $0.1836 / 1M tokens |
| Output | $0.7128 / 1M tokens |
| Release date | 2025-04-05github.com/meta-llama ↗ | verified |
| Architecture | Auto-regressive Mixture-of-Experts model with early-fusion multimodality; 16 expertsgithub.com/meta-llama ↗ | verified |
| Active parameters | 17B per forward passgithub.com/meta-llama ↗ | verified |
| Total parameters | 109Bgithub.com/meta-llama ↗ | verified |
| Context window | 10,000,000 tokensgithub.com/meta-llama ↗ | verified |
| Input and output modalities | Multilingual text and image input → multilingual text and code outputgithub.com/meta-llama ↗ | verified |
| Training token count | ~40 trillion tokensgithub.com/meta-llama ↗ | verified |
| Knowledge cutoff | August 2024github.com/meta-llama ↗ | verified |
| Supported languages | Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, and Vietnamesegithub.com/meta-llama ↗ | verified |
| License | Llama 4 Community Licensegithub.com/meta-llama ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU Pre-trained model evaluation; 5-shot macro_avg/acc_char reported by Meta for Llama 4 Scout. | 79.6%% | github.com/meta-llama ↗ | |
MMLU Pro Instruction-tuned model evaluation; 0-shot macro_avg/acc reported by Meta. | 74.3%% | github.com/meta-llama ↗ | |
GPQA Diamond Instruction-tuned model evaluation; 0-shot accuracy reported by Meta. | 57.2%% | github.com/meta-llama ↗ | |
MATH Pre-trained model evaluation; 4-shot em_maj1@1 reported by Meta. | 50.3%% | github.com/meta-llama ↗ | |
MMMU Instruction-tuned image reasoning evaluation; 0-shot accuracy reported by Meta. | 69.4%% | github.com/meta-llama ↗ | |
MathVista Instruction-tuned image reasoning evaluation; 0-shot accuracy reported by Meta. | 70.7%% | github.com/meta-llama ↗ | |
ChartQA Instruction-tuned image understanding evaluation; 0-shot relaxed_accuracy reported by Meta. | 88.8%% | github.com/meta-llama ↗ | |
MBPP Pre-trained code evaluation; 3-shot pass@1 reported by Meta. | 67.8%% | github.com/meta-llama ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "meta/llama-4-scout",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat endpoint. Switch only the baseURL and model name — no SDK changes required.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/llama-4-scout",
"messages": [{"role": "user", "content": "Summarize this document."}]
}'Meta launched the Llama 4 family on April 5, 2025, introducing Scout (17Bx16E, 109B total, 10M context) and Maverick (17Bx128E, 400B total, 1M context). Both use early-fusion MoE architectures for native image and text reasoning, with Scout positioned around ultra-long context and Maverick around stronger per-token capability.
Scout and Maverick both activate 17B parameters per forward pass and support multimodal input. Scout has 16 experts, 109B total parameters, and a 10M token context window. Maverick has 128 experts, 400B total parameters, and a 1M token context window. Scout is the longer-context route; Maverick is the stronger benchmark route.
Yes. Meta publishes Llama 4 Scout as open weights under the Llama 4 Community License. Meta's model card says Scout is released as BF16 weights and can fit within a single H100 GPU with on-the-fly int4 quantization, but production deployments still need their own safety, evaluation, and serving controls.
Meta's Llama 4 model card says Llama 4 has been tested for image understanding up to 5 input images. Treat anything beyond that as unvalidated and test it in your own application before relying on it.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Architecture | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Active parameters | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Total parameters | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Context window | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Input and output modalities | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Training token count | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Knowledge cutoff | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Supported languages | github.com/meta-llama ↗ | 2026-08-25 | verified |
| License | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MMLU | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MMLU Pro | github.com/meta-llama ↗ | 2026-08-25 | verified |
| GPQA Diamond | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MATH | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MMMU | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MathVista | github.com/meta-llama ↗ | 2026-08-25 | verified |
| ChartQA | github.com/meta-llama ↗ | 2026-08-25 | verified |
| MBPP | github.com/meta-llama ↗ | 2026-08-25 | verified |
| Meta releases Llama 4 Maverick and Scout — open multimodal MoE herd | ai.meta.com ↗ | 2026-08-25 | verified |
| How is Llama 4 Scout different from Llama 4 Maverick? | github.com/meta-llama ↗ | 2026-08-25 | to verify |
| Can I self-host Llama 4 Scout? | github.com/meta-llama ↗ | 2026-08-25 | to verify |
| How many images can Llama 4 Scout handle per request? | github.com/meta-llama ↗ | 2026-08-25 | to verify |