TranslateGemma 27B

Overview

TranslateGemma 27B is the largest size in Google's TranslateGemma family and the one Google describes as built for maximum fidelity. Google fine-tuned the family from Gemma 3 checkpoints: the card lists 4.3 billion tokens for supervised fine-tuning, mostly translation pairs written by humans or generated by Gemini, followed by a reinforcement learning phase of 10.2 million tokens.

Google's WMT24++ results (55 languages) put this size first in the family, with MetricX at 3.09 (lower is better) and COMET at 84.4. Across the 10 WMT25 pairs that Google had professional translators score, it reached 5.86 MQM against 7.94 for the 12B model, and it beat the 12B on every one of those pairs.

Choosing a variant: TranslateGemma 12B trails it by about half a MetricX point on WMT24++ and suits everyday text. TranslateGemma 4B handles short, informal content. TranslateGemma 4B Image translates text found in pictures.

What you can build

• Drafts of articles, manuals and policy pages, translated section by section for a reviewer to check.
• App store listings, release notes and in-product copy for many locales from one source text.
• A reference translation for comparing the output of smaller models or human translators.
• Translation steps in agent and content pipelines where quality matters more than volume.

Inputs

The source text goes into the prompt field. Source and target languages come from two menus of 54 languages each. The default run translates a Chinese passage about ancient beliefs on death into English.

Maximum New Tokens, under the advanced settings, accepts 1 to 2,048 and starts at 200.

According to Google's card, every size works within a total input context of 2K tokens, so a long document needs to be translated in parts.

Outputs

One text file with the translation in the target language.

Recommended settings

Set Maximum New Tokens high enough to hold the whole expected translation, up to 2,048. The English version of the default Chinese passage needs more than 200 tokens, so at the default it ends partway through.

There is no field for tone or style instructions. The text is wrapped in the fixed translation prompt Google trained the model with, so put only the text you want translated in the prompt field.

Limitations

• In Google's human evaluation, Japanese to English was the one pair where TranslateGemma 27B scored clearly worse than the base Gemma 3 27B, which the report traces to mistranslated named entities. Check names in the output.
• With German as the target, it scored about the same as base Gemma 3 27B in that evaluation.
• WMT24++, the benchmark behind Google's 55-language numbers, only tests translation out of English. Google published results for just three other directions (Czech to Ukrainian, Czech to German and Japanese to English), all in its human evaluation.

Safety & compliance

TranslateGemma 27B is covered by the Gemma Terms of Use and the Gemma Prohibited Use Policy. Google's safety testing (child safety, harassment, hate speech, violence, bias and stereotyping) ran without safety filters and found minimal policy violations, with improvements over earlier Gemma models on ungrounded inferences. Have a qualified human translator review sensitive or regulated content before it is published.