**Google Just Unveiled TranslateGemma: A New Family of Open Translation Models for 55 Languages**
Hey there, language enthusiasts!
In a big move, Google AI has announced TranslateGemma, a new family of open machine translation models designed to support an impressive 55 languages. These models are built on top of Gemma 3, a large language model developed by Google, and come in 4B, 12B, and 27B parameter sizes.
**What’s the Deal with TranslateGemma?**
So, what exactly is TranslateGemma? It’s not a separate model, but rather a specialized version of Gemma 3 that’s been fine-tuned for translation using a two-stage pipeline. The pipeline consists of:
1. **Supervised fine-tuning on large parallel corpora**, which involves combining human translations with high-quality artificial translations generated by Gemini models.
2. **Reinforcement learning** that optimizes translation quality using a multi-sign reward ensemble.
**Getting the Data Right**
The supervised fine-tuning stage starts from the public Gemma 3 4B, 12B, and 27B checkpoints. Researchers used parallel data that combines human translations with artificial translations generated by Gemini models. They also used human-generated parallel data from the SMOL and GATITOS datasets, which cover 123 and 170 languages, respectively. This improved representation of scripts and language families that are underrepresented in publicly available Web parallel data.
**Reinforcement Learning: The Next Step**
After supervised fine-tuning, TranslateGemma runs a reinforcement learning stage on the same translation data combination. This stage uses a multi-sign reward ensemble that includes:
* MetricX 24 XXL QE, which approximates MQM scores and is used in quality estimation mode without a reference.
* Gemma AutoMQM QE, a span-stage error predictor fine-tuned from Gemma 3 27B IT on MQM-labeled data.
* ChrF, a character n-gram overlap metric that compares model output with artificial references.
* A Naturalness Autorater that uses the coverage model as an LLM selector and produces span-stage penalties for segments that don’t sound like native text.
* A generalist reward model from the Gemma 3 post-training setup that retains reasoning and instruction-following capacity.
**Benchmark Results: TranslateGemma Takes the Lead**
The TranslateGemma models were evaluated on the WMT24++ benchmark using MetricX 24 and Comet22. The results show that TranslateGemma improves quality across all 55 language pairs, with the 12B model surpassing the 27B Gemma 3 baseline and the 4B model reaching quality similar to the 12B baseline.
**Multimodal Translation: A New Frontier**
TranslateGemma inherits the picture understanding stack of Gemma 3 and is evaluated on the Vistra benchmark. The model receives only the picture and a prompt that asks it to translate the text in the picture, without explicit OCR or bounding box input. The results show that TranslateGemma improves MetricX from 2.03 to 1.58 and Comet22 from 76.1 to 77.7.
**Key Takeaways**
* TranslateGemma is a specialized Gemma 3 variant designed for translation that supports 55 languages.
* The model combines Gemini artificial data with human parallel corpora to improve representation of both high-resource and low-resource languages.
* The reinforcement learning stage uses an ensemble of quality estimation rewards to optimize translation quality.
* Smaller models can match or beat larger Gemma 3 baselines on WMT24++ across 55 languages.
* The models retain multimodal capabilities and are launched as open weights on Hugging Face and Vertex AI.
Check out the full paper on the arXiv website, and look for the model weights on Hugging Face and Vertex AI. Join the conversation on Twitter and follow us on Reddit and our Newsletter for more machine learning news and updates!
