The technical map of the Llama infrastructure at Meta
Meta builds the Llama infrastructure and publishes its model weights, itself a significant technical commitment. But the licence is called the Llama 4 Community License and is not a standard open source licence: it is free up to a ceiling of 700 million monthly active users, and above that ceiling a separate licence from Meta is required.
- Weights are published
- A custom licence, not standard open source
- Persian is not on the language support list
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Model families
The Llama 4 line, two members with two different expert-layout architectures.
Versions
| Versions | Deployed | Status | Model families |
|---|---|---|---|
| Llama 4 Maverick | current | Llama 4 | |
| Llama 4 Scout | current | Llama 4 |
Persian is not among the documented languages
This is the most critical parameter in this report for a Persian-speaking reader. On the model card for both Llama 4 members, Meta logs exactly twelve languages: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. Persian is not logged on that list.
The structure of this list matters, because it differs from the Mistral case. Meta does not use the phrase "including"; it provides a closed list of twelve. So the absence of Persian here is not an indirect inference — it is a documented parameter. The model can likely produce Persian text, but Meta has not claimed its quality and has not logged official support for it, and for a workload that must be delivered to a client, these two parameters are not equivalent.
The licence is not standard open source, and that needs precise documentation
"Llama is open source" is a claim repeated everywhere. The actual licence, the Llama 4 Community License, carries three conditions a standard open source licence lacks. First, if on the release date of that version you had more than 700 million monthly active users, a separate licence from Meta is required. Second, "Built with Llama" must be displayed in a visible position. Third, the name of any derived model must begin with "Llama."
For nearly every reader of this report, the first condition is inert and the other two are operationally active. Fine-tune the model, apply a different name to it, and the licence is in breach. The licence text also states that use must comply with trade regulations including export controls, and this report neither advises on that point nor describes a route around it.
Larger model, smaller window
This parameter is counterintuitive and is visible in the table. Scout carries 109 billion total parameters and a ten-million-token context window. Maverick carries 400 billion total parameters and a one-million-token window, a tenth of Scout. Both activate 17 billion parameters per pass; what differs is the expert count, 16 against 128.
So for processing one very long document, the smaller model is the correct choice. The assumption that a larger figure always guarantees better performance is wrong in this case.
Knowledge cutoff, the oldest parameter in this reference
Both Llama 4 models are logged at August 2024. For comparison, Opus 5 is logged at May 2026, roughly twenty-one months later. For workloads dependent on recent libraries or recent events, this is Llama 4 largest operational limitation, and it has no bearing on the model processing quality.
A small parameter that costs real time
The Llama documentation has moved from llama.com to developer.meta.com, and the old address redirects with a 301. Top-level links function; deeper paths do not necessarily.
Documented strengths
- Weights are genuinely published, so the model runs on the user own hardware infrastructure with no dependency on any account
- Scout ten-million-token window, the highest figure any maker in this reference has documented
- Both models are multimodal and accept image as well as text input
- The model card logs total parameters, active parameters and expert count separately, a level of transparency not every maker provides
Known weaknesses and limits
- Persian is not among the twelve languages Meta lists. Arabic is logged.
- The knowledge cutoff is August 2024, the oldest parameter among the live models in this reference.
- The licence is not standard open source: it carries a user ceiling, mandatory "Built with Llama" display, and a requirement that derived model names begin with Llama.
- Meta publishes no API rate for these models, because the operational product is the weights, not a service. The user operating cost equals hardware cost.
Technical verdict
For a workload that must run on the user own infrastructure and processes very long documents, Scout with its ten-million-token window is a serious proposal. For a Persian-language workload, this family is not the first option and is not recommended in this report.
Frequently raised questions, with documented answers
Does Llama process Persian
Meta logs twelve languages on the Llama 4 model card and Persian is not among them, meaning Meta has not claimed Persian processing quality. The model may produce Persian text, but that is not equivalent to documented official support.
Is Llama genuinely open source
The weights are published but the licence is not standard open source. It is the Llama 4 Community License, carrying a 700-million monthly active user ceiling, mandatory "Built with Llama" display, and a requirement that derived model names begin with Llama.
Scout or Maverick
Scout carries a ten-million-token window and Maverick one million, even though Maverick is the larger model. For long documents, Scout is the correct choice, and both activate 17 billion parameters per pass.