What is Llama 4 Scout?
Llama 4 Scout is an open-weights model from Meta, released on 5 April 2025 with Llama 4 Maverick. It is now an older line: Meta replaced Llama with its Muse models in April 2026, has released no Llama model since April 2025, and its own Model API lists only Muse models. Llama 4 Behemoth, the nearly two-trillion-parameter teacher Meta previewed at launch, never shipped.
Scout is a mixture-of-experts model with 109 billion parameters in total and 17 billion active per token, spread over 16 experts. It reads text and up to five tested images and writes text and code, in 12 supported languages. Meta advertises a 10-million-token context, though it was trained at 256K and extended in mid-training; the knowledge cutoff is August 2024.
The weights are gated on Hugging Face under the Llama 4 Community License, which is not open source. Hosts include DeepInfra at $0.10 input and $0.30 output per million tokens and Google Vertex AI at $0.25 and $0.70. Groq, once its fastest host, retired it for free and developer tiers on 17 July 2026.
How much does Llama 4 Scout cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Llama 4 Community License, gated on Hugging Face |
| DeepInfra | $0.10 in / $0.30 out per 1M tokens | FP8; $0.08 / $0.24 on the flex tier |
| Google Vertex AI | $0.25 in / $0.70 out per 1M tokens | 8,192-token output cap via OpenRouter |
Free weights under the Llama 4 Community License (gated). Hosted: DeepInfra $0.10 input / $0.30 output per 1M tokens ($0.08 / $0.24 on its flex tier); Google Vertex AI $0.25 / $0.70.
Compare this price with other models
Is Llama 4 Scout worth it?
Scout's selling point was a 10-million-token context, and independent tests did not bear it out. On Fiction.LiveBench at 120,000 tokens it scored 15.6%, against 28.1% for Maverick and 90.6% for Gemini 2.5 Pro, as reported by The Decoder. Meta trained it at 256K and stretched it further in mid-training; hosts sell much less, with DeepInfra listing 327,680 tokens on OpenRouter and Google Vertex AI 1,310,720.
On short tasks it is a capable mid-sized model by 2025 standards. Meta's own card puts it at 74.3 on MMLU Pro, 57.2 on GPQA Diamond and 94.4 on DocVQA, all self-reported on BF16 weights. Artificial Analysis first scored it 36 on its Intelligence Index, about level with GPT-4o mini, and 43 after a grading change. That is well behind what open models of its size reach in late 2026.
The licence carries real conditions. Companies with more than 700 million monthly active users on 5 April 2025 must ask Meta for a licence, products must show "Built with Llama", and fine-tuned models must have names starting with "Llama". Meta's acceptable use policy also withholds the rights for its multimodal models from individuals and companies based in the European Union, though their end users are not affected; the clause is still on llama.com.
Hosting is shrinking. Groq shut Scout off for free and developer tiers on 17 July 2026 and pointed users to gpt-oss-120b, and Meta's Model API never carried it. Meta's own newer open model, Muse Glimmer, is a 30B dense model under Apache 2.0 with none of the Llama terms. Scout's remaining argument is price: at DeepInfra's $0.10 and $0.30 it costs a third to a quarter of Glimmer's $0.30 and $1.20 on the same host.
Where it earns its place
An e-commerce data team tagging 2 million product photos
Write short attribute lists and captions from each image and title
Send each photo and title to Scout on DeepInfra; at about 1,500 input and 150 output tokens per item the batch costs roughly $300 input and $90 output. The 5-image testing limit and short prompts suit this better than long documents.
A platform engineer with one 80GB H100 and a Llama 3 deployment
Move to a newer Llama with image input without buying more GPUs
Meta says Scout fits on one H100 with on-the-fly int4 quantisation, and Unsloth's 4-bit GGUF builds are about 58 to 65GB. Keep prompts well under 128K tokens, since long-context accuracy drops sharply.
Before you commit
- Setup
- Accept Meta's licence on Hugging Face to download meta-llama/Llama-4-Scout-17B-16E-Instruct, then serve it with vLLM, Transformers or llama.cpp; or call it through OpenRouter, DeepInfra or Vertex AI with no setup.
- Integrations
- OpenAI-compatible APIs at DeepInfra and through OpenRouter, Google Vertex AI Model Garden; tool calling supported on most listed endpoints; GGUF builds for llama.cpp, Ollama and LM Studio.
- Data and privacy
- Self-hosted, nothing leaves your servers. The Llama 4 Community License requires a Meta licence above 700M monthly users, "Built with Llama" attribution and "Llama"-prefixed names for derived models, and its use policy excludes EU-based companies from the multimodal rights.
- Cost at scale
- 1B input and 250M output tokens a month is about $175 on DeepInfra. Self-hosting needs about 217GB in BF16 (at least three 80GB GPUs), about 115GB at 8-bit and about 58 to 65GB at 4-bit, which fits one 80GB card.
- Limits
- Advertised 10M-token context, trained at 256K, weak beyond 100K in independent tests; tested with up to 5 images; 12 supported languages; static model with an August 2024 cutoff and no planned updates.
What’s good
- $0.10 input and $0.30 output per 1M tokens on DeepInfra
- Reads images natively; 94.4 on DocVQA (Meta, self-reported)
- Only 17B active parameters, so it is fast for its 109B size
- Fits on one 80GB GPU at 4-bit, per Meta
What’s not
- 15.6% on Fiction.LiveBench at 120K tokens, far short of the 10M claim
- Llama 4 licence: 700M-MAU clause, naming rules and an EU multimodal exclusion
- No new Llama release since April 2025; Groq retired it in July 2026
- Knowledge cutoff of August 2024
Best for
Teams already running Llama 4 who need cheap hosted text and image handling outside the EU and can live with an unsupported model line.
Llama 4 Scout alternatives
Llama 4 Scout: common questions
- Is Llama 4 Scout worth it?
- Scout is still one of the cheapest vision-capable models to rent, but its headline context window was never real and Meta itself has left the Llama line behind. We rate it 3.2 out of 5.
- How much does Llama 4 Scout cost?
- Free weights under the Llama 4 Community License (gated). Hosted: DeepInfra $0.10 input / $0.30 output per 1M tokens ($0.08 / $0.24 on its flex tier); Google Vertex AI $0.25 / $0.70.
- What are the downsides of Llama 4 Scout?
- 15.6% on Fiction.LiveBench at 120K tokens, far short of the 10M claim; llama 4 licence: 700M-MAU clause, naming rules and an EU multimodal exclusion; no new Llama release since April 2025; Groq retired it in July 2026; knowledge cutoff of August 2024.
- Who is Llama 4 Scout best for?
- Teams already running Llama 4 who need cheap hosted text and image handling outside the EU and can live with an unsupported model line.
- How much does Llama 4 Scout cost for a team or in production?
- 1B input and 250M output tokens a month is about $175 on DeepInfra. Self-hosting needs about 217GB in BF16 (at least three 80GB GPUs), about 115GB at 8-bit and about 58 to 65GB at 4-bit, which fits one 80GB card.
- Is Llama 4 Scout safe for company data?
- Self-hosted, nothing leaves your servers. The Llama 4 Community License requires a Meta licence above 700M monthly users, "Built with Llama" attribution and "Llama"-prefixed names for derived models, and its use policy excludes EU-based companies from the multimodal rights.
- What are the best Llama 4 Scout alternatives?
- The closest alternatives we have reviewed are Muse Glimmer, Gemma 4, gpt-oss-120b and Qwen3.8-27B.
Answers checked Nov 9, 2026.