TranslateGemma 12B Image

Overview

This is the 12B size of TranslateGemma applied to images. You give it a picture, say which language is written on it and which one you want, and it returns the translated text. Google made TranslateGemma by fine-tuning Gemma 3 checkpoints on 4.3 billion tokens of translation data (publicly available parallel documents, plus web documents paired with Gemini translations), followed by a reinforcement learning phase on 10.2 million tokens.

Google's headline result for this size is on text: measured with MetricX on WMT24++, TranslateGemma 12B outperforms the original Gemma 3 27B while using less than half the parameters. The card lists 3.60 MetricX and 83.5 Comet on WMT24++, 7.94 MQM on WMT25, and 2.08 MetricX on Vistra, its image translation test. On Vistra the lead does not hold: Google's technical report puts Gemma 3 27B at 2.03 MetricX and 76.1 Comet, ahead of this model's 2.08 and 72.8.

Choosing a variant: TranslateGemma 27B Image uses an identical form and does better on all four of those numbers, for example 5.86 against 7.94 MQM on WMT25. 12B makes sense when a working translation is enough, for instance to search or sort images by what they say. If the output goes straight to readers, the 27B scores make the stronger case.

What you can build

• Expense apps that read foreign receipts and invoices.
• Language-learning apps where a student photographs a worksheet and reads it in their own language.
• Moderation workflows that need a working English version of the text inside user-uploaded images.

Inputs

• Input Image: a JPG, PNG, WebP, GIF or HEIC file, uploaded or referenced by URL.
• Source Language and Target Language: two dropdowns that share a list of 54 languages, including Hindi, Thai, Hebrew, Persian and Irish. The target defaults to English.
• Maximum New Tokens: how much translated text may come back, 1 to 2048, default 200.

Outputs

Plain translated text, delivered as a .txt file. Positions of the text in the image are not returned.

Recommended settings

• Short items like a sign or a label fit within the default 200 tokens. Raise the cap toward 2048 for multi-paragraph pages.
• A wide banner or a long strip gets scaled into one 896 x 896 square before the model reads it. Cutting it into a few roughly square crops keeps more pixels per letter.
• Set Source Language to the language actually written in the image. TranslateGemma's prompt format passes a source language code with every request.

Limitations

• Check names and key terms. This page's sample is a Greek passage in a decorative typeface, run with Source Language left at its Bulgarian default instead of Greek. The output gets the setting, Scythia, right, but it renders the name Hephaestus as Hecate, and much of the rest does not match the original.
• Google notes that the model is not a knowledge base and can produce incorrect or outdated statements.

Safety & compliance

The Gemma terms of use and the Gemma Prohibited Use Policy cover this model. Google trained it for one stated purpose, translating text or image input, and makes no claims about other abilities. The card encourages developers to add content safety safeguards that fit their own product policies, and says the training data was filtered to remove certain personal information.