Applied artificial intelligence, neural systems, and smart agents.
Fusing generative AI, multi-agent frameworks, and high-throughput data processing to automate business workflows.
Llama 4 Scout
Scout is specified with a ten million token context window — the largest figure any maker publishes in this reference and ten times its nearest rival. Its architecture carries 109 billion total parameters and 17 billion active across 16 experts, and its weights are released, though the operating licence is not standard open source.
- Ten million token window capacity
- Publicly released weights
- Outside the Persian coverage boundary
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Persian sits outside the specified language list
The most load-bearing line in this document for a Persian-speaking deployment. Meta model card lists exactly twelve languages as official coverage: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. Persian is not recorded on that list.
The list is closed and is not defined as extensible with an including clause, so the absence of Persian is not an estimate — it is a specification boundary. The system will likely still produce Persian output, but Meta has not certified its quality in the official specification. That may be sufficient for internal workloads; for anything routed into a client deliverable, the gap between produces and is supported is exactly what surfaces as cost later.
What the ten million token capacity is actually provisioned for
The largest window capacity on any other model in this reference is one million tokens. Scout is ten times that, and this is the one component of the architecture where Llama 4 genuinely leads this reference.
The capacity earns its keep in loading one very long document in a single pass: a full code repository, a legal case file, an archive. It has no practical use in ordinary conversation, where the large window only raises memory consumption without changing answer quality.
Which parameter figure to read
109 billion total parameters, 17 billion active across 16 experts. As with any expert-based architecture, the active figure sets the compute cost per token and the total figure sets the memory footprint required for deployment. On that measure Scout is far lighter than Mistral Large 3, which carries 675 billion total parameters.
The licence architecture is not standard open source
Llama is open source is repeated everywhere. The actual specification is the Llama 4 Community License, and it embeds three conditions a standard open source licence does not carry: a ceiling of 700 million monthly active users, mandatory prominent display of a Built with Llama notice, and a requirement that any derived model name begin with Llama. The first is inert for most deployments; the other two are real and enforceable.
The oldest knowledge cutoff in this reference
August 2024. For comparison, Opus 5 is set at May 2026, roughly twenty one months ahead. If a workload is tied to recent libraries or recent events, this is an important area for operational consideration when recent data is required, and it has nothing to do with processing quality.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- not checked
- Payment
- not checked
- Free tier
- no
How we checked: No hosted cloud service at a published price point exists for this model; Meta output is the weight set itself, not a hosted endpoint. So the question of reachability from Iran shifts into a different problem here: deployment on your own infrastructure. This is derived from the licence text, not from a direct network probe, and the licence itself places an explicit compliance requirement — including export controls — into the user obligation set.
Documented strengths
- A ten million token window capacity, the largest figure recorded in this reference and ten times the nearest rival
- Publicly released weights, so deployment runs on your own infrastructure and depends on no account
- 17 billion active parameters, keeping compute cost per token at a low tier
- A multimodal architecture, accepting image alongside text as input
Room for improvement and technical considerations
- Persian is not on the twelve-language specification list Meta publishes; Arabic is.
- The knowledge cutoff is August 2024, the oldest figure among the live models in this reference.
- The licence architecture is not standard open source: a user ceiling is set, Built with Llama display is mandatory, and derived model names must begin with Llama.
- Meta has not published a price point for this model, because the product output is the weights themselves. Cost here is infrastructure cost.
Technical verdict
If the workload requires reading one very long document in a single pass and deployment runs on your own infrastructure, Scout is the best option in this reference today. If the workload is Persian-centric or tied to recent information, look elsewhere.
Frequently raised questions, with documented answers
Does Llama 4 Scout cover Persian
Meta model card lists twelve languages and Persian is not one of them, so Meta has not certified Persian output quality in the official specification. The system may still produce Persian, but that is not the same as declared support.
Does a ten million token window capacity always mean better performance
No. The capacity only carries operational value when the input text is genuinely long. In ordinary conversation, the large window only raises memory consumption without changing answer quality.
Scout or Maverick, which to provision
Scout carries the ten million token window capacity and 109 billion total parameters; Maverick carries 400 billion total parameters but a one million token window capacity. For long documents, Scout is the better fit. Both activate 17 billion parameters per pass.