Skip to content

DirectoryTools

TranslateGemma logo

TranslateGemma

Google's open translation models in 4B, 12B and 27B for 55 languages.

Visit TranslateGemma

TranslationFreeWe checked the prices on huggingface.co on 28 Sep 2026.

TranslateGemma homepage screenshot

Recaply review

TranslateGemma review: free offline translation in 55 languages

TranslateGemma is Google's family of open translation models in 4B, 12B and 27B sizes, built on Gemma 3 for developers who want to translate text and images on their own hardware.

$0
To download all three sizes from Hugging Face or Kaggle
55
Languages Google trained and tested it on, plus about 500 more language pairs in training
12B
Model beats Gemma 3 27B at translation on WMT24++, by Google's MetricX test
2K
Tokens of input per request, either text or one image

What is TranslateGemma?

TranslateGemma comes from Google's Translate research team: Gemma 3 models trained on human and Gemini-made translations, then tuned further with reinforcement learning. The 4B is meant for phones and edge devices, the 12B for laptops and the 27B for a single H100 GPU or TPU. Each one translates a passage of text, or the text in a photo, between two language codes you name.

It is still Google's current translation model; Google has released no successor, and the newer Gemma 4 is a general-purpose family. Unlike DeepL or Google Translate, you host it yourself and pay nothing per character.

Should you use TranslateGemma or Gemma 4 for translation?

Start with TranslateGemma for plain translation jobs, and try Gemma 4 when you need long inputs, instructions or speech.

TranslateGemma is Gemma 3 retrained only for translation, and Google publishes translation scores for it. Gemma 4 is Google's newer general model family: it reads 128K to 256K tokens, follows system prompts, takes audio on its smaller sizes and uses the Apache 2.0 licence, but its model card lists no translation benchmark. Some local users say Gemma 4 31B translates better and others prefer TranslateGemma, so run both on your own text.

Who is TranslateGemma best for?

TranslateGemma is best for developers who want free translation that runs offline or on their own servers, in apps, on devices or as a base to fine-tune for one language pair.

  • Offline and on-device apps

    The 4B is sized for phones and edge devices, so translations can run with no network connection and no text leaving the device.

  • Private translation at volume

    Hosting it on your own GPU means no per-character fees and no text sent to an outside service.

  • Fine-tuning for one language pair

    Google built it as a starting point for training your own model for a specific language pair or a low-resource language.

  • Signs, menus and scanned pages

    Send a photo and it reads the text in it and returns the translation, an image skill it keeps from Gemma 3.

What's included with TranslateGemma?

Three sizes from phone to single GPU, translation of the text inside images, 55 tested languages with regional codes, and deployment from Google Cloud's Model Garden.

Three sizes
4B for phones and edge devices, 12B for consumer laptops and 27B for one H100 GPU or TPU in the cloud.
Image translation
Send an image instead of text and it pulls out the words and translates them; images are resized to 896 x 896 pixels.
55 tested languages
Google tested it on 55 languages, from Spanish and Hindi to many low-resource ones, and trained it on nearly 500 more language pairs without published scores.
Regional language codes
Source and target take ISO language codes, with regional variants such as en-GB or pt-BR.
Google Cloud deployment
Deploy it from Google Cloud's Model Garden to an endpoint in your own project, with a ready-made notebook.

How much does TranslateGemma cost?

TranslateGemma is free: Google gives away all three sizes as open weights on Hugging Face and Kaggle, and there is no paid plan or hosted per-token price.

Open weights$0Free download
  • 4B, 12B and 27B on Hugging Face and Kaggle
  • Accept Google's Gemma terms to download
  • You pay only for the hardware you run it on

Google sells no hosted TranslateGemma API with a per-token price: Hugging Face lists no inference provider for it, and Google Cloud's Model Garden deploys it to an endpoint in your own Google Cloud project. The weights come under the Gemma Terms of Use, which have no non-commercial clause. Full pricing on huggingface.co. We checked these prices on .

How do readers rate TranslateGemma?

No one has rated TranslateGemma yet. Be the first.

No ratings yetThe score and its stars show once 3 readers have rated TranslateGemma.
Your rating

Sign in to rate. One rating per account.

Your rating of TranslateGemma

TranslateGemma pros and cons

It is free, runs offline and translates well for its size, but it takes only 2K tokens at a time and some users get better results from Gemma 4.

Pros

  • Free to download and run offline; users run the 4B on an 8 GB phone and call it an offline Google Translate.
  • Strong for its size: Google's tests put the 12B ahead of Gemma 3 27B, and a translation agency's subtitle test ranked it above several hosted models on automatic scores.
  • Developers see it as a good base for fine-tuning on one language pair, which is how Google designed it.
  • Machine translation users on Reddit single out its wide language list, including many less common languages.

Cons

  • Takes only 2K tokens of input per request, so long documents have to be split into pieces.
  • Some local users find the larger Gemma 4 models translate better, especially Arabic and Korean.
  • Users report weak Traditional Chinese output, which testers trace to its training data.
  • You must name the source language on every request, and it only translates: there is no chat, instructions or other task.

What are the best TranslateGemma alternatives?

Three tools on Recaply do TranslateGemma's job. Gemma 4 is open source, so you can run it on your own server.

Questions about TranslateGemma

What is TranslateGemma?

TranslateGemma is Google's family of open translation models in 4B, 12B and 27B sizes, built on Gemma 3 for developers who want to translate text and images on their own hardware.

Who is TranslateGemma for?

TranslateGemma is best for developers who want free translation that runs offline or on their own servers, in apps, on devices or as a base to fine-tune for one language pair.

Is TranslateGemma free?

Yes. TranslateGemma has a free Open weights plan: 4B, 12B and 27B on Hugging Face and Kaggle.

Is there an open source alternative to TranslateGemma?

Yes. Gemma 4 is an open source tool that does the same job, so you can run your own.

What are the best TranslateGemma alternatives?

The closest alternatives on Recaply are Gemma 4, Tiny Aya and DeepL.

Where can I try TranslateGemma?

On its official website, blog.google.

Can I use TranslateGemma commercially?

Yes, within Google's Gemma Terms of Use, which have no non-commercial clause. You must follow the Gemma prohibited use policy, and if you share the model or a modified version you must pass on the terms and a notice file. Google claims no rights in the translations.

Does TranslateGemma work offline?

Yes. Once you download the weights, TranslateGemma runs on your own hardware with no connection: Google sizes the 4B for phones and edge devices and the 12B for consumer laptops, and pitches the 27B at a single H100 GPU or TPU in the cloud.

How much text can TranslateGemma translate at once?

TranslateGemma takes 2K tokens of input per request, so split long documents into smaller pieces and send them one by one. An image counts as 256 tokens after it is resized to 896 x 896 pixels, and you send one image per request.

Which languages does TranslateGemma support?

TranslateGemma was trained and tested on 55 languages, from Spanish, French, Chinese and Hindi to many low-resource ones. Google also trained it on nearly 500 more language pairs but has not published quality scores for them, so test those pairs before you rely on it.

Does TranslateGemma have an API?

Not a pay-per-use one from Google. You download the weights and run them yourself, or deploy them from Google Cloud's Model Garden to an endpoint in your own project; Hugging Face lists no hosted inference provider for it.

Our verdict

TranslateGemma is a strong free pick when you need private, offline translation and can host a model yourself.

Google's tests show the 12B beating a Gemma 3 model twice its size, users run the 4B on phones, and the Gemma terms have no non-commercial limit.

Skip it for
Long documents or instruction-style tasks, since it reads 2K tokens at a time and only translates.
Check first
Test your own language pairs against Gemma 4, which some users find translates better.