Translation

Translation models turn text in one language into another without sending it off the device. Two kinds are available behind the same API:

• Opus-MT: small (~110 MB) and fast, one model per language pair, such as English to German.
• EuroLLM-1.7B-Instruct: one 1.7 GB model that translates between any two of its 35 languages.
• Hy-MT2-1.8B: one 1.2 GB model that translates between any two of its 38 languages and dialects, and is stronger than EuroLLM on Chinese.

Quick Start

The useTranslator hook downloads the .pte model and tokenizer, loads the pipeline, and manages its lifecycle. Every call names its language pair as ISO 639-1 codes:

import { models, useTranslator } from 'react-native-executorch';

function MyComponent() {
  const translator = useTranslator(models.translation.OPUS_MT_EN_DE.DEFAULT);

  // Hook state:
  // translator.isReady          — true once model and tokenizer are loaded
  // translator.downloadProgress — 0 to 100 download progress
  // translator.error            — Error instance if download or load failed

  const handleTranslate = async (text: string) => {
    if (!translator.isReady || !translator.translate) return;

    const german = await translator.translate(text, { sourceLanguage: 'en', targetLanguage: 'de' });
    console.log(german);
  };
}

An Opus-MT model accepts only its own pair. With models.translation.EUROLLM_1_7B.DEFAULT or models.translation.HY_MT2_1_8B.DEFAULT, the pair can change on every call without reloading the model.

Streaming

Pass a callback to receive the translation as it is decoded. Each call gets the newly decoded piece of text; joined together, the pieces equal the returned translation.

let partial = '';
const result = await translator.translate('How are you? I am fine.', languages, (piece) => {
  partial += piece;
  setText(partial);
});

Call translateStop() to cancel; translate() then resolves with the text translated so far.

How Input Is Processed

Line breaks are kept in the output for both kinds of model.

• Opus-MT models are trained on single sentences, so the input is split after sentence-ending punctuation (., !, ?) and at line breaks, and each sentence is translated on its own. A single sentence longer than the model's maximum source length (256 tokens) is truncated, and one sentence translates to at most 256 tokens.
• EuroLLM and Hy-MT2 translate each line on their own with the model card's prompt, decoding greedily. A line and its translation together must fit the 1024-token context.

EuroLLM supports Arabic (ar), Bulgarian (bg), Catalan (ca), Chinese (zh), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), Galician (gl), German (de), Greek (el), Hindi (hi), Hungarian (hu), Irish (ga), Italian (it), Japanese (ja), Korean (ko), Latvian (lv), Lithuanian (lt), Maltese (mt), Norwegian (no), Polish (pl), Portuguese (pt), Romanian (ro), Russian (ru), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Turkish (tr) and Ukrainian (uk).

Hy-MT2 supports Arabic (ar), Bengali (bn), Burmese (my), Cantonese (yue), Chinese (zh), Traditional Chinese (zh-Hant), Czech (cs), Dutch (nl), English (en), Filipino (tl), French (fr), German (de), Gujarati (gu), Hebrew (he), Hindi (hi), Indonesian (id), Italian (it), Japanese (ja), Kazakh (kk), Khmer (km), Korean (ko), Malay (ms), Marathi (mr), Mongolian (mn), Persian (fa), Polish (pl), Portuguese (pt), Russian (ru), Spanish (es), Tamil (ta), Telugu (te), Thai (th), Tibetan (bo), Turkish (tr), Ukrainian (uk), Urdu (ur), Uyghur (ug) and Vietnamese (vi).

Imperative API

For use outside React components, create the pipeline with createMarianTranslator or createLLMTranslator:

import { createLLMTranslator, download, models } from 'react-native-executorch';

const model = await download(models.translation.EUROLLM_1_7B.DEFAULT);
const translator = await createLLMTranslator(model);

try {
  console.log(
    await translator.translate('The weather is nice today.', {
      sourceLanguage: 'en',
      targetLanguage: 'pl',
    })
  );
} finally {
  translator.dispose();
}

translateWorklet runs the Opus-MT pipeline synchronously on a worklet runtime. See Worklets & Threading.

Available Models

The models are available in models.translation:

Model	Direction	Size	Supported Backends
OPUS_MT_EN_DE	English → German	107 MB	XNNPACK (CPU)
OPUS_MT_DE_EN	German → English	107 MB	XNNPACK (CPU)
OPUS_MT_EN_ES	English → Spanish	114 MB	XNNPACK (CPU)
OPUS_MT_ES_EN	Spanish → English	114 MB	XNNPACK (CPU)
OPUS_MT_EN_FR	English → French	108 MB	XNNPACK (CPU)
OPUS_MT_FR_EN	French → English	108 MB	XNNPACK (CPU)
EUROLLM_1_7B	Any of its 35 languages	1.7 GB	XNNPACK (CPU), MLX (iOS, GPU)
HY_MT2_1_8B	Any of its 38 languages	1.2 GB (MLX 1.8 GB)	XNNPACK (CPU), MLX (iOS, GPU)

License

OPUS_MT_EN_DE is licensed under CC-BY 4.0 and requires attribution to the Helsinki-NLP Opus-MT project. The other Opus-MT pairs, EuroLLM and Hy-MT2 are Apache-2.0.

API Reference

Hooks & Pipelines

• useTranslator() — React hook for translation model downloading, state, and lifecycle.
• createMarianTranslator() — Imperative factory for Opus-MT translation pipelines.
• createLLMTranslator() — Imperative factory for EuroLLM and Hy-MT2 translation pipelines.

Types

• nlp.TranslationLanguages — Source and target language of one call.
• TranslatorModel — Any translation model configuration.
• MarianTranslator — Opus-MT runner, which adds translateWorklet.
• MarianTranslatorModel — Opus-MT configuration with model path, tokenizer path, and language pair.
• LLMTranslator — EuroLLM and Hy-MT2 runner.
• LLMTranslatorModel — EuroLLM or Hy-MT2 configuration with model, tokenizer, and tokenizer config paths and the translation prompt.

Model Presets

• models.translation — Pre-configured translation models registry.

Source Code

View the implementation on GitHub:

• src/extensions/nlp/tasks/marianTranslation.ts ↗
• src/extensions/llm/tasks/llmTranslation.ts ↗