MarianMT — small, fast, one model per language pair

Why it matters

When you know the exact directions your app needs, per-pair models are the efficient choice: ship only those pairs, load them fast, and keep the footprint small — all on-device, no server, no per-call fees.

• Fast per-pair localization — bundle EN→DE/FR/ES for your target markets without a 470 MB download.
• Offline, private translation — translate messages or docs with nothing leaving the tab.
• Round-trip QA — pair with a reverse model to sanity-check machine translations.
• Pipeline stage — feed a classifier or TTS a translated string, all locally.

Use cases

How the API works

Each pair is its own model; the pipeline call needs no language codes:

import { pipeline } from "@huggingface/transformers";

const translate = await pipeline(
  "translation",
  "Xenova/opus-mt-en-de",
); // downloads + caches ~105 MB once — this pair only

const out = await translate("Hello, how are you?");
// → [{ translation_text: "Hallo, wie geht es dir?" }]

There's no src_lang/tgt_lang — the direction is baked into the model. To add a pair, load its repo (opus-mt-en-fr, opus-mt-en-es, …). This page streams tokens via a TextStreamer inside a Web Worker so the UI stays responsive.

References

• Model card — MarianMT EN→DE (Xenova ONNX)
• Original weights — Helsinki-NLP/opus-mt-en-de
• OPUS-MT — the Helsinki-NLP project
• Marian NMT — the framework the models are named for
• Transformers.js documentation