TranslateGemma
Google's Gemma 3-based translation models in 4B, 12B and 27B sizes for 55 languages, which also translate text inside images.
In short. TranslateGemma is a set of AI translation models from Google. You pick a size: 4B for phones and small devices, 12B for a laptop and 27B for a single data-center GPU. They translate between 55 languages and can also read the text in a photo and translate it. The files are free to download, but you have to accept Google's Gemma terms first.
Summary
Google announced TranslateGemma on 15 January 2026. It is a set of three translation models fine-tuned from Gemma 3, called 4B, 12B and 27B, with about 4.30, 12.19 and 27.43 billion parameters. Google evaluated them on 55 language pairs and trained them on nearly 500 more pairs that it has not evaluated; the list is in the technical report as a starting point for research. The input is a text string or an image, which is normalized to 896 × 896 pixels and encoded to 256 tokens, and the output is the translated text. The model card gives a total input context of 2K tokens, so longer documents have to be split.
Google trained the models in two stages. Supervised fine-tuning used 4.3 billion tokens of human translations and synthetic translations written by Gemini, and reinforcement learning then used 10.2 million tokens, with reward models that include MetricX-QE and AutoMQM. Google says the 12B model beats the Gemma 3 27B baseline on the WMT24++ benchmark as measured by MetricX, and that the 4B model rivals the 12B baseline. The model card lists MetricX scores of 5.32, 3.60 and 3.09 for the 4B, 12B and 27B models on WMT24++ (lower is better) and COMET scores of 81.6, 83.5 and 84.4. On a WMT25 test in 10 languages the 12B and 27B models score 7.94 and 5.86 on the MQM scale, where lower is better. Google also reports gains on Vistra, a benchmark of text in images, although it did not fine-tune the models for images. These are all Google's own evaluations.
In BF16 the weights are 8.6 GB, 24.4 GB and 54.9 GB. Google positions the 4B model for mobile and edge devices, the 12B model for consumer laptops and the 27B model for a single H100 GPU or TPU in the cloud. The chat template is strict. It supports only the user and assistant roles, and the user message must be one entry with the type text or image, a source language code and a target language code, either plain, such as en, or regional, such as en-GB. An unsupported code raises an error, and other prompt styles such as post-editing are not officially supported. The Hugging Face repositories are public but gated: you accept Google's usage license, and requests are processed immediately. The license is the Gemma Terms of Use, not the Apache 2.0 license of Gemma 4, because TranslateGemma is built on Gemma 3. On 11 November 2026 Hugging Face showed about 1.1 million downloads of the 4B model and about 960,000 of the 12B model since release.
Good for
- Self-hosted translation between 55 languages on a laptop with the 12B model, on one GPU with the 27B model or on small devices with the 4B model
- Translating text in photos, signs and screenshots with the same model
- A starting point for fine-tuning extra or low-resource language pairs, since Google also trained on nearly 500 pairs it has not evaluated
- Pipelines that pass explicit source and target language codes, including regional variants such as en-GB
Watch out for
- The license is the Gemma Terms of Use, not Apache 2.0, so its Prohibited Use Policy must be passed on to anyone you distribute the model to
- The model card gives a total input context of 2K tokens, so long documents have to be split, and free-form instructions or glossaries are not part of the documented template
- It only translates, and Google makes no claim about other abilities, so it is not a chat model
- The quality scores are Google's own tests on WMT24++, WMT25 and Vistra, so evaluate on your own text, and note that Google did not fine-tune it for images
- Language codes that the model does not support raise an error instead of falling back to a guess
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the TranslateGemma 4B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| TranslateGemma 4B BF16 | 4.3B | 9.89 GB | 4.94 GB | 2.97 GB | Laptop or phone |
| TranslateGemma 12B BF16 | 12.2B | 28 GB | 14 GB | 8.41 GB | Laptop or phone |
| TranslateGemma 27B BF16 | 27.4B | 63 GB | 32 GB | 19 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
Google's custom license for the Gemma models released before Gemma 4, including TranslateGemma. It allows commercial use, changes and redistribution with no revenue cap, but every copy and every derivative must carry the same use restrictions, and Google may restrict use it believes breaks the terms.
- Attribution or model name must be shown
- Acceptable-use policy applies
See also
- Index-Translate vs TranslateGemma: which one to pick?
- Best open-source translation models in 2026
- Four open translation models, four different yardsticks
- Index-Translate: Bilibili's open translators for 150 languages
- PLaMo 3 Translate: PFN's open translators for 53 languages
Frequently asked questions
- What is TranslateGemma?
- TranslateGemma is a set of AI translation models from Google. You pick a size: 4B for phones and small devices, 12B for a laptop and 27B for a single data-center GPU. They translate between 55 languages and can also read the text in a photo and translate it. The files are free to download, but you have to accept Google's Gemma terms first.
- Can I use TranslateGemma commercially?
- Yes, with conditions. TranslateGemma is released under the Gemma Terms, which allows commercial use but adds extra rules: Attribution or model name must be shown; Acceptable-use policy applies. Read the license text before you ship a product.
- Is TranslateGemma open source?
- TranslateGemma is open-weight, not open source in the strict sense. You can download the model files, but the Gemma Terms adds conditions that the Open Source Initiative definition does not allow.
- Can I run TranslateGemma on my own computer?
- Yes. The smallest variant, TranslateGemma 4B, needs about 2.97 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was TranslateGemma released?
- TranslateGemma was released by Google DeepMind on January 15, 2026.
- How long a context window does TranslateGemma have?
- TranslateGemma supports a context window of up to 2K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- TranslateGemma 12B model card (Hugging Face) huggingface.co
- TranslateGemma: A new suite of open translation models (Google blog, 15 January 2026) blog.google
- TranslateGemma Technical Report (arXiv 2601.09012) arxiv.org
- Gemma Terms of Use ai.google.dev
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Hy-MT2 Tencent | 2.04B – 30.1B 3B active | 256K | Apache 2.0 (Permissive) | 1.41 GB Laptop or phone | |
| Index-Translate Bilibili | 2.27B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 1.44 GB Laptop or phone | |
| PLaMo 3 Translate Preferred Networks | 890M – 31.8B | 32K | PLaMo Community (Conditional) | 0.614 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |