TranslateGemma 27B Image
Google's largest TranslateGemma model, used here to read the text in a picture and return it in another language. Its card gives it the family's best scores on both text and image translation tests.
Overview
TranslateGemma 27B Image, the largest TranslateGemma model, reads the text in a picture and writes it out in the language you choose. Google built the family from Gemma 3 checkpoints, fine-tuned them on parallel text that mixes human translations with translations generated by Gemini, and then added a reinforcement learning stage.
Image reading comes from the Gemma 3 base. On the Vistra benchmark, Google found that the text translation gains carried over to text inside images, even without image-specific fine-tuning.
Google calls this size "built for maximum fidelity", and the card lists the family's best results for it. On WMT24++, which covers 55 languages, it reaches 3.09 MetricX and 84.4 Comet. It scores 5.86 MQM on WMT25 and 1.57 MetricX on Vistra, where lower is better for both metrics.
Choosing a variant: TranslateGemma 12B Image offers the same inputs, posts 2.08 on Vistra against this model's 1.57 and trails it on the text tests too. Pick 27B when people will read the translation as final text, such as a document you publish or a reply shown to a customer.
What you can build
- Translation of scanned book pages and letters. This page's sample is the opening of the Gospel of John in Turkish, returned in English with the verse numbers kept.
- Help desks that translate a customer's screenshot of an error dialog.
- Travel apps that translate a photographed street sign or menu.
Inputs
- Input Image: the picture with the text, uploaded or added by URL, in JPG, PNG, WebP, GIF or HEIC.
- Source Language: the language of the text in the picture. It defaults to Turkish to match the sample.
- Target Language: defaults to English. Both lists hold the same 54 languages, Swahili, Urdu, Basque and Welsh among them.
- Maximum New Tokens: an upper limit on the length of the translation, from 1 to 2048, default 200.
Per Google's card, each image is resized to 896 x 896 pixels and encoded as 256 tokens, and the whole input is limited to 2K tokens.
Outputs
A text file with the translation. In the sample, "Sevindirici Haber" and "Yuhanna" become "THE GOOD NEWS" and "JOHN", followed by verses 1 to 8 in English.
Recommended settings
- The sample's eight verses fit inside the default 200 tokens. For a full page of body text, raise the limit, or the translation stops partway through.
- Crop to the part of the photo that holds the text, since small print in a large or very wide shot loses detail at 896 x 896.
Limitations
- Google's Vistra score uses images filtered to contain a single text, so multi-column pages and tables are outside what it measured.
- The card warns that the model can miss subtle nuance, sarcasm and figurative language.
- Only text comes back. The translation is not placed onto the original image.
Safety & compliance
Google releases TranslateGemma under its Gemma terms of use, and prohibited uses are listed in the Gemma Prohibited Use Policy. The stated intended use is translation from text or image input. Google's safety testing, run without safety filters, found minimal policy violations for image-to-text at every size.
API quick start
Run translate-gemma-27b-it-image with a single API call.
{
"inputImage": "https://your-cdn.com/input.png",
"sourceLng": "Turkish",
"targetLng": "English",
"maxNewTokens": 200
}curl -X POST "https://api.wiro.ai/v1/Run/google/translate-gemma-27b-it-image" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"inputImage": "https://your-cdn.com/input.png",
"sourceLng": "Turkish",
"targetLng": "English",
"maxNewTokens": 200
}
JSONOther models you might like
Explore similar wiro ai models.
