> ## Documentation Index
> Fetch the complete documentation index at: https://api-docs.ollang.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Llama for Translation

> Meta Llama 4 in Ollang — natively multimodal mixture-of-experts models with a 10M-token context window and 12 officially supported languages.

**Llama 4** is Meta's first natively multimodal, mixture-of-experts open-weight generation. Both variants run **17 billion active parameters** and accept multilingual text and images, producing multilingual text and code.

<Info>
  **Available in Ollang Workflows** as `llama` — Document Translation and Subtitle Translation. See the [Translation catalog](/providers/translation).
</Info>

<Note>
  Llama 4 does not serve AI Dubbing Orders. Select a different provider for the `translate` step of a dubbing Workflow.
</Note>

## Models

| | Llama 4 Scout | Llama 4 Maverick |
| - | - | - |
| **Active parameters** | 17B | 17B |
| **Total parameters** | 109B | 400B |
| **Experts** | 16 | 128 |
| **Context window** | 10M tokens | 1M tokens |
| **Training tokens** | \~40 trillion | \~22 trillion |
| **Positioned for** | Single-GPU efficiency, extreme context | Image and text understanding, fast responses |

Knowledge cutoff for both is **August 2024** — the oldest in this catalog, which matters for content containing recent product names, terminology, or events.

## Supported languages

Meta officially supports **12 languages**: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, and Vietnamese.

This is the single most important constraint to design around. Outside those 12, quality falls off sharply compared with the 100+ language providers.

## Key capabilities

* **Extreme context length** — Scout's 10M-token window is by a wide margin the largest here, enough for entire books, full documentation sets, or complete season subtitle runs in a single pass.
* **Native multimodality** for documents combining text and images.
* **Open weights**, relevant for organizations running parallel self-hosted evaluation or fine-tuning for domain-specific glossaries.

## Where it fits in a workflow

Llama 4 is the right choice for **high-volume work inside its 12 supported languages** — bulk subtitle backlogs, internal documentation, and draft passes destined for human review. For the same content in an unsupported language, [Agentic Turbo](/apis/translation-apis/ollang-agentic-localization) or [DeepL](/apis/translation-apis/deepl) will produce better output.

## Trade-offs

* **Narrow official language coverage** — 12 languages.
* **Oldest knowledge cutoff in the catalog** at August 2024.
* **No AI Dubbing support.**
* Draft-grade rather than deliverable-grade on nuance-heavy content; budget for review.

## Reference

* [Translation provider catalog](/providers/translation)
* [Llama 4 Scout model card](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.