Opus-MT
Free.ai (self-hosted)
·
translation
·
~100 tokens per translation
Opus-MT is a translation model built by Helsinki-NLP. Strongest at Commercial translation including low-resource languages.. Self-hosted on Free.ai GPUs — runs free against your daily token pool (100 tokens per message). Released under Apache 2.0 — commercial use permitted on Free.ai.
Use via API
OpenAI-compatible REST API. Generate a key and call this model in seconds.
curl -X POST https://api.free.ai/v1/translate/ \
-H "Authorization: Bearer sk-free-..." \
-H "Content-Type: application/json" \
-d '{"model":"opus-mt","text":"...","target":"es"}'Frequently Asked Questions
Opus-MT supports hundreds of language pairs. MadLAD-400 covers 419 languages; NLLB covers 200. Auto-detect handles the source language; pick the target from a searchable list.
Modern multilingual models (especially MadLAD-400) are notably stronger on low-resource languages than older NMT systems. Quality is still uneven — pair with a native speaker review for production publishing in rare languages.
Tags and markdown are preserved when you check the "preserve formatting" toggle. Inline code blocks stay literal. For document translation see /pdf/translate/.
Yes — prefix names with `[[KEEP:name]]` or use the glossary feature on the API. Brand names, product names, and people names stay literal across the output.
Opus-MT runs on our GPUs and is among the cheapest tools — ~100 tokens per call. Comes out of your daily free pool first; $5 = 200K paid tokens.
5,000 characters per request on the web UI. For longer documents use /pdf/translate/ which chunks and stitches automatically, or call the API in a loop.
Yes — output is correctly RTL-oriented. Mixed RTL/LTR text inside one paragraph is also handled (e.g., embedded English brand names in Arabic prose).
Yes — POST a list of strings to /v1/translate/. Each translation in the same order as input. /batch/ also accepts a CSV.
Yes — POST to /v1/translate/ with model="Opus-MT" plus source + target ISO codes. /api/ has full reference + auth + error handling.
Self-hosted models keep your text on our GPUs and discard it after the call returns. Premium pass through with a DPA. We do not train on your inputs.
Yes — Free.ai grants commercial use of translated output for websites, documents, support tickets, marketing copy, software localization.
Sub-second for short text on self-hosted. Long passages (>1,000 words) take 2–5 seconds. The API streams output token-by-token if you set stream=true.