Llama 4 Scout
Scout has a ten million token context window, the largest figure any maker publishes in this reference and ten times its nearest rival. It has 109 billion total parameters and 17 billion active across 16 experts, and its weights are published, though the licence is not standard open source.
- Ten million token window
- Published weights
- Persian is not listed
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Persian is not among the listed languages
The most important sentence on this page for a reader here. On the model card Meta lists exactly twelve languages: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. Persian is not one of them.
Note that the list is closed and does not say including, so the absence is not a guess. The model will probably produce Persian, but Meta has not claimed its quality. For internal work that may be enough; for something you hand to a client, the difference between those two is what costs you later.
What ten million tokens is actually for
The largest window on any other model in this reference is one million. Scout is ten times that, and this is the one place where Llama 4 genuinely comes first here.
It earns its keep where one very long document has to be read in one go: a whole code repository, a legal case file, an archive. Where it does not earn its keep is ordinary conversation, where a large window only consumes more memory without making the answer better.
Which parameter number to read
109 billion total, 17 billion active across 16 experts. As with any expert-based model, the active figure tells you the compute cost per token and the total tells you how much memory you need. On that measure Scout is far lighter than Mistral Large 3, which carries 675 billion total.
The licence is not open source
Llama is open source is repeated everywhere. The actual licence is the Llama 4 Community License and it carries three conditions a standard open source licence does not: a ceiling of 700 million monthly active users, mandatory prominent display of Built with Llama, and a requirement that any derived model name begin with Llama. The first is meaningless for you and the other two are real.
The oldest knowledge cutoff in this reference
August 2024. For comparison, Opus 5 sits on May 2026, about twenty one months ahead. If your work is tied to recent libraries or recent events, this is the model biggest practical weakness and it has nothing to do with its 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: Meta does not sell a hosted service at a published price for this model; the product is the weights themselves. So the question of whether it opens from Iran takes a different shape here: a model with published weights runs on your own hardware. We read that in the licence rather than measuring it, and the licence itself says your use must comply with trade compliance laws including export controls.
What it is good at
- A ten million token window, the largest figure in this reference and ten times the next one
- Published weights, so it runs on your own hardware and depends on no account
- 17 billion active parameters, so the compute cost per token is low
- Multimodal, taking image input as well as text
Where it falls short
- Persian is not among the twelve languages Meta lists. Arabic is.
- The knowledge cutoff is August 2024, the oldest among the live models in this reference.
- The licence is not standard open source: a user ceiling, mandatory Built with Llama display, and derived names must begin with Llama.
- Meta publishes no price for this model, because the product is the weights. Your cost is hardware.
Our take
If you have to read one very long document in a single pass and you run the model yourself, Scout is the best option in this reference today. If your work is in Persian or tied to recent information, look elsewhere.
Questions people actually ask
Does Llama 4 Scout know Persian
Meta lists twelve languages on the model card and Persian is not one of them, so Meta has not claimed Persian quality. The model may well produce Persian, but that is not the same as declared support.
Does a ten million token window mean it is always better
No. A large window is only worth something when you actually feed the model long text. For ordinary conversation it just consumes more memory without improving the answer.
Scout or Maverick
Scout has the ten million window and 109 billion total parameters; Maverick has 400 billion total but a one million window. For long documents, Scout. Both activate 17 billion parameters per pass.