How to choose an open-source translation model

Compare open translation models by language coverage, license, size and instruction support, and find which one fits documents, subtitles, Japanese, images or a laptop.

  • Beginner
  • 6 min read
  • Published
  • Updated

Short answer. For self-hosted text translation, start with Hy-MT2 or Index-Translate, which are both plain Apache 2.0, and add TranslateGemma if you can accept Google’s Gemma terms. Pick PLaMo 3 Translate mainly for Japanese, and read its revenue-limited license first. Choose by license, language list and the size you can run rather than by headline scores, because every score is the developer’s own, and test two or three candidates on your own text.

What a translation model does today

Machine translation used to mean a system built only for translating. The open models in this guide are different: each starts from a general language model, such as Gemma 3 or Qwen3.5, and is trained further on translation data. The result is a model that translates well at a small size, and that usually does nothing else.

Developers say the specialist training pays off. Google reports that TranslateGemma 12B beats the Gemma 3 27B baseline, a model more than twice its size, as measured by MetricX on WMT24++. Preferred Networks reports that its 0.8B PLaMo 3 Translate beats larger models in Japanese-English tests. Both are the developers’ own numbers, but they explain why a 4B to 12B specialist is a reasonable first choice. The translation models list shows everything we track, including image and speech models.

The shortlist

Model Sizes Languages License Released What stands out
Hy-MT2 1.8B, 7B, 30B-A3B 33 Apache 2.0 May 2026 Templates that keep tags and structured data; 256K context; a 440 MB on-device build
Index-Translate 2B, 9B, 35B-A3B (preview) 150 Apache 2.0 Sep 2026 Glossary, format and tone controls; translates text in images
TranslateGemma 4B, 12B, 27B 55 Gemma Terms of Use Jan 2026 Translates text in images; strict language-code template; 2K input
PLaMo 3 Translate 0.8B, 2B, 8B, 31B 53 PLaMo Community Nov 2026 Tuned for Japanese; glossary and polite or plain Japanese; 0.57 GB GGUF

Two other groups of models translate too. HunyuanOCR finds text in menus, posters and screenshots and translates it. Three speech models translate audio: Granite Speech 5.0 to and from English in six languages, Nemotron 3.5 Transcribe to or from English in 31 languages, and LFM3-Audio across 18 languages on a device. How to choose a speech-to-text model covers the audio side.

Pick by what you translate

  • Documents and web content in common languages. Hy-MT2 or Index-Translate. Check that your language is on the list first, because the lists differ: 33 languages for Hy-MT2, plus Tibetan, Kazakh, Mongolian, Uyghur and Cantonese on its language table, and 150 for Index-Translate.
  • Text with tags, placeholders, JSON or HTML. Hy-MT2 provides templates that preserve delimiters and structured data. Index-Translate supports hard constraints that keep JSON, CSV, Markdown, code blocks and placeholders unchanged.
  • A fixed glossary. Index-Translate and PLaMo 3 Translate accept a glossary, and Hy-MT2 has a terminology template. The TranslateGemma template has no glossary field.
  • Japanese. PLaMo 3 Translate, which also lets you choose plain or polite Japanese and pass context. PLaMo 3 Translate vs Hy-MT2 lays out the trade-off.
  • Low-resource languages. Index-Translate reports a dedicated low-resource test across 62 languages, and the Hy-MT2 language table also lists Tibetan, Kazakh, Mongolian, Uyghur and Cantonese. Treat both as claims to verify.
  • Text inside photos, signs and screenshots. TranslateGemma and Index-Translate both accept images, and HunyuanOCR reads the text first. Index-Translate vs TranslateGemma compares the two translators.
  • A phone, laptop or CPU. The 0.8B and 2B models: PLaMo 3 Translate 0.8B and 2B, Index-Translate-2B and Hy-MT2-1.8B.

Check the license first

Hy-MT2 and Index-Translate use Apache 2.0, which allows commercial use with few conditions. The other two have more:

  • TranslateGemma uses the Gemma Terms of Use. You can use it commercially, but any agreement you use to share it must carry Google’s Prohibited Use Policy, models distilled from its outputs inherit the terms, and the Hugging Face download needs you to accept the license. Gemma 4 is Apache 2.0, but TranslateGemma is built on Gemma 3, so the older terms apply.
  • PLaMo 3 Translate uses the PLaMo Community License. It is free for personal and academic use, and free for commercial use after registration while annual revenue including affiliates stays at or below 1 billion yen. Above that you need a separate agreement, and derived models must carry “PLaMo” in their name.

AI licenses explained covers how to read these terms.

Do not rank them by their headline scores

Each developer’s report puts its own model ahead. Preferred Networks’ charts place PLaMo 3 Translate ahead of TranslateGemma and Hy-MT2. The Index-Translate report places its 9B model ahead of both on eight of nine columns, and Google compares TranslateGemma only with its Gemma 3 baseline. We found no independent comparison of the four.

The numbers also use different scales and tests. TranslateGemma’s card gives COMET as 83.5 on a 0 to 100 scale for 12B on WMT24++ across 55 languages, while the Index-Translate report gives 0.8601 on a 0 to 1 scale for its 9B model, with English into 20 languages. The glossary entry on translation metrics explains BLEU, chrF++, COMET and MetricX, and the Benchmark entry explains why self-reported scores need care. Use the scores to build a shortlist, not to pick a winner.

Test on your own text

Pick 50 to 100 sentences that look like your real input, in your own language pairs, and run two or three candidates on them. Then check:

  1. Names, numbers, dates and units. A wrong figure is the error you can least afford.
  2. Terminology. Does the model keep your glossary terms, and does it translate product names it should leave alone?
  3. Tags and placeholders, such as HTML, {variable} or JSON keys.
  4. The right language. A model can answer in a language other than the one you asked for, which the model cards call an off-target output.
  5. Dropped or invented content, especially in long paragraphs.
  6. Tone and politeness, which matter a lot in Japanese, Korean and German.

Ask a speaker of the target language to read the results. How to evaluate AI models yourself shows how to turn this into a repeatable test set.

Memory and setup

By our estimate a 7B to 9B model needs about 5.5 to 6.7 GB of memory at 4-bit and about 16 to 19 GB in BF16, and a 27B to 31B model needs about 19 to 22 GB at 4-bit. A mixture-of-experts model such as Hy-MT2-30B-A3B or the Index-Translate 35B-A3B preview still has to hold all its weights, so plan for 21 to 25 GB at 4-bit. You can try your own numbers in the VRAM calculator, and quantization explains the 4-bit figures.

Setup differs more than quality does:

  • Hy-MT2 needs Transformers 5.6 or newer with trust_remote_code, and has GGUF builds for llama.cpp.
  • Index-Translate serves through vLLM with a build that supports Qwen3.5. Its prompts are written in Chinese, and image translation in GGUF needs a separate mmproj file.
  • TranslateGemma loads through the Transformers image-text-to-text pipeline with a strict chat template. Each message carries one entry with a source and a target language code, and a code the model does not support raises an error.
  • PLaMo 3 Translate is meant to run through the vLLM Completions API with a hand-built prompt for production, and through the Transformers chat template, which needs custom code, for trials. Its input and output together are limited to 32,768 tokens.

TranslateGemma accepts only 2K tokens of input, so long documents have to be split in every case, and a model with a short context window loses earlier sentences. Once a model fits your memory and license, GGUF files make local trials simple through llama.cpp.

Frequently asked questions

What is the best open-source translation model?
There is no single best model. Hy-MT2 and Index-Translate are plain Apache 2.0 and are good first candidates, with Index-Translate listing 150 languages against 33 for Hy-MT2. TranslateGemma and PLaMo 3 Translate come with extra license conditions. Every quality claim is the developer's own, so translate a sample of your own text with two or three of them.
Can I use these translation models commercially?
Yes, with conditions on two of them. Hy-MT2 and Index-Translate are Apache 2.0. TranslateGemma uses the Gemma Terms of Use, which allow commercial use but require you to pass its use policy on. PLaMo 3 Translate is free for commercial use after registration while your revenue stays at or below 1 billion yen. HunyuanOCR, which also translates text in images, is not licensed in the EU, the UK or South Korea.
Which model is best for Japanese?
PLaMo 3 Translate was tuned and measured mostly on Japanese and English and offers plain or polite Japanese as an option, but those results are Preferred Networks' own and the license has a revenue limit. Hy-MT2 and Index-Translate also cover Japanese, so test all three on your own text.
Can I run a translation model on a laptop?
Yes. PLaMo 3 Translate 0.8B has a 0.57 GB GGUF file, Index-Translate-2B a 1.31 GB one, and Tencent offers a 440 MB 1.25-bit build of the 1.8B Hy-MT2. By our estimate the 7B to 9B models need about 5.5 to 6.7 GB of memory at 4-bit, which suits a laptop with 16 GB.
Can a general chat model translate instead?
Often it can, and it is enough when you only need the gist. Google says TranslateGemma 12B beats the larger Gemma 3 27B baseline on WMT24++ as measured by MetricX, which suggests that specialist training helps. The specialists also offer language-code templates, glossaries and tag handling, but they only translate and cannot chat.
How do I know whether a translation is good?
Have a speaker of the target language read 50 to 100 sentences of your own text. Check names, numbers, terms, tags and tone, and look for sentences that came out in the wrong language or lost content. Scores such as COMET and MetricX help a developer compare versions but are not comparable between model cards.

Key terms in this guide

  • Benchmark A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read.
  • Context window The context window is the maximum amount of text, measured in tokens, that a model can take into account at one time, including your prompt and its own answer.
  • GGUF GGUF is a single-file model format used by llama.cpp and the tools built on it, which bundles quantized weights with the information needed to run them.
  • Machine translation Machine translation is software that translates text, speech or text inside images from one language to another. The open models of today usually start from a general model and are then trained only to translate.
  • Quantization Quantization stores a model's weights with fewer bits, for example 4 instead of 16, to cut memory use and speed up inference at a small cost in quality.
  • Self-hosting Self-hosting means running a model on hardware you control, such as your own PC or a rented GPU server, instead of calling a vendor's API.
  • Translation metrics Translation metrics are the scores that model cards use to rate a translation. Some count overlap with a human reference, others use a trained model or human reviewers, and their scales cannot be compared.

Models mentioned in this guide

Model Released Parameters Context License Smallest variant needs
Hy-MT2 Tencent 2.04B – 30.1B 3B active 256K Apache 2.0 (Permissive) 1.41 GB Laptop or phone
Index-Translate Bilibili 2.27B – 36B 3B active 256K Apache 2.0 (Permissive) 1.44 GB Laptop or phone
TranslateGemma Google DeepMind 4.3B – 27.4B 2K Gemma Terms (Conditional) 2.97 GB Laptop or phone
PLaMo 3 Translate Preferred Networks 890M – 31.8B 32K PLaMo Community (Conditional) 0.614 GB Laptop or phone
HunyuanOCR Tencent 1.12B 128K Tencent Hunyuan (Conditional) 0.773 GB Laptop or phone
Granite Speech 5.0 IBM 470M – 2.31B – Apache 2.0 (Permissive) 1.09 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
LFM3-Audio-2B Liquid AI 2B 32K LFM Open 1.0 (Conditional) 4.6 GB Laptop or phone