TranslateGemma 4B

Overview

TranslateGemma 4B is the smallest of Google's three open translation models built on Gemma 3. Google trained the family in two stages: supervised fine-tuning on parallel text that mixes human translations with translations generated by Gemini, then a reinforcement learning phase guided by an ensemble of reward models that includes MetricX-QE and AutoMQM. In use it does one job. You give it text, a source language and a target language, and it returns the translation.

Google built the 4B size with mobile and edge deployment in mind and reports that it rivals the larger Gemma 3 12B baseline on the WMT24++ benchmark. The model card lists a MetricX score of 5.32 for this size across the 55 WMT24++ languages, where lower is better.

Choosing a variant: TranslateGemma 12B and TranslateGemma 27B score better on the same benchmarks and suit text where exact wording matters, such as published copy. TranslateGemma 4B Image runs this same 4B model on photos and screenshots instead of typed text.

What you can build

• Translation of short user-generated content: reviews, chat messages, support tickets, product questions.
• First-pass localisation of UI strings and help-center snippets ahead of a human review.

Inputs

• The text to translate, pasted into the prompt box.
• A source language and a target language, each picked from a menu of 54 languages. It covers widely spoken languages such as English, Chinese, Spanish, Arabic and Hindi, and lower-resource ones including Swahili, Estonian, Lithuanian and Marathi.
• Maximum New Tokens (advanced), from 1 to 2,048 with a default of 200. It caps the length of the translation.

Google's card lists a total input context of 2K tokens, so long documents should go in as separate sections.

Outputs

Plain text in the target language, returned as a text file. The prompt Google ships with the model asks for the translation alone, with no added commentary.

Recommended settings

The 200-token default fits a few sentences. At that setting, a long paragraph such as the ten-sentence Chinese passage this page starts with is cut off before its last sentences. For full paragraphs, raise Maximum New Tokens well above the default. The cap only limits length, so short input still gets a short answer.

Limitations

• Google's card says longer context generally gives better output, up to a point, so a lone word, name or label is harder to get right than a full sentence.
• The card also warns that subtle nuance, sarcasm and figurative language can be missed.
• Quality varies by language. In Google's per-language results, lower-resource targets such as Estonian, Lithuanian and Swahili come out much less accurate than Dutch or Vietnamese, and the gap is widest at this size.
• At 4B parameters it trails the 12B and 27B sizes on every benchmark in Google's card, and it was not part of Google's human evaluation.

Safety & compliance

TranslateGemma is released under the Gemma Terms of Use, and use must follow Google's Gemma Prohibited Use Policy. Google's evaluation covered child safety, content safety and representational harms, and the card reports minimal policy violations at all sizes. Google asks developers to add content safeguards that fit their own product, so place any checks before or after the translation step.