TranslateGemma 12B
The mid-sized model in Google's TranslateGemma family: Gemma 3 fine-tuned for text translation across 55 languages. Google reports it outperforms the larger Gemma 3 27B baseline on WMT24++.
Overview
TranslateGemma 12B is the middle size of Google's TranslateGemma family, a set of open models trained from Gemma 3 for translation. It takes a passage, a source language and a target language, and returns the translated text.
Google's headline result for the family concerns this size. On WMT24++, a benchmark covering 55 languages, TranslateGemma 12B outperforms the Gemma 3 27B baseline on MetricX while using less than half the parameters. The card gives it a MetricX score of 3.60 (lower is better) and a COMET score of 83.5 on that set. Professional translators also rated it on 10 language pairs from WMT25, where it reached an MQM score of 7.94, again lower being better. Google designed the 12B model to run on consumer laptops.
Choosing a variant: TranslateGemma 4B is the lighter option for short, high-volume text. TranslateGemma 27B scores highest in the family and fits text headed for publication. For writing inside images, use TranslateGemma 4B Image.
What you can build
- Localisation of product descriptions, listings and landing pages for several markets at once.
- Translation of documentation, knowledge-base articles and onboarding emails, section by section.
- Multilingual customer support that translates both the customer's message and the agent's reply.
- Translated datasets or subtitle drafts for a human editor to finish.
Inputs
You paste the source text into the prompt field and choose two languages from menus with 54 options, among them English, French, Russian, Japanese, Turkish, Vietnamese, Persian and Catalan. The default example on this page turns an English scene description into Russian.
An advanced field, Maximum New Tokens, limits the length of the reply. It starts at 200 and accepts values from 1 to 2,048.
The model card lists a total input context of 2K tokens. Split long files at paragraph or section breaks and send each piece as its own request.
Outputs
The translation as plain text, delivered as a text file.
Recommended settings
If a translation ends mid-sentence, the reply has hit the token limit. Raise Maximum New Tokens and run it again.
Send whole sentences or paragraphs rather than fragments. Google's card says output quality depends on the amount of context provided, and a full paragraph gives the model more to work with than a lone heading.
Limitations
- The lead over Gemma 3 27B comes from automatic metrics. In Google's human evaluation on 10 WMT25 language pairs, base Gemma 3 27B scored better on 7 of them, and the 12B model matched or beat it only on English to Italian, Marathi and Swahili.
- Google notes that biases or gaps in the training data can show up as limitations in the output.
- Google built and evaluated the model for translation. Other tasks, such as summarising or rewriting tone, fall outside its intended use.
Safety & compliance
Use falls under the Gemma Terms of Use and Google's Gemma Prohibited Use Policy. Google red-teamed the models without safety filters for child safety, harmful content and bias, and reports minimal policy violations at every size. A translation reflects its source text, so filter content in your own pipeline where that matters.
API quick start
Run translate-gemma-12b-it with a single API call.
{
"prompt": "A person stands in profile beneath an orn…",
"sourceLng": "English",
"targetLng": "Russian",
"maxNewTokens": 200
}curl -X POST "https://api.wiro.ai/v1/Run/google/translate-gemma-12b-it" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A person stands in profile beneath an orn…",
"sourceLng": "English",
"targetLng": "Russian",
"maxNewTokens": 200
}
JSONOther models you might like
Explore similar wiro ai models.
