TranslateGemma Technical Report

Abstract: We present TranslateGemma, a suite of open machine translation models based on the Gemma 3 foundation models. To enhance the inherent multilingual capabilities of Gemma 3 for the translation task, we employ a two-stage fine-tuning process. First, supervised fine-tuning is performed using a rich mixture of high-quality large-scale synthetic parallel data generated via state-of-the-art models and human-translated parallel data. This is followed by a reinforcement learning phase, where we optimize translation quality using an ensemble of reward models, including MetricX-QE and AutoMQM, targeting translation quality. We demonstrate the effectiveness of TranslateGemma with human evaluation on the WMT25 test set across 10 language pairs and with automatic evaluation on the WMT24++ benchmark across 55 language pairs. Automatic metrics show consistent and substantial gains over the baseline Gemma 3 models across all sizes. Notably, smaller TranslateGemma models often achieve performance comparable to larger baseline models, offering improved efficiency. We also show that TranslateGemma models retain strong multimodal capabilities, with enhanced performance on the Vistra image translation benchmark. The release of the open TranslateGemma models aims to provide the research community with powerful and adaptable tools for machine translation.

• The paper demonstrates enhanced translation quality by integrating supervised fine-tuning with reinforcement learning using composite reward signals across 55 language pairs.
• The methodology employs a two-phase training approach with curated synthetic and human-authored data, ensuring robust performance in both text and image-based translation.
• Evaluation reveals significant reductions in MetricX and Comet22 scores, outperforming baseline models in both high-resource and low-resource settings.

TranslateGemma: Architecture, Training Paradigm, and Evaluation for Open Multilingual Machine Translation

Introduction

TranslateGemma is a suite of open-source machine translation (MT) models derived from the Gemma 3 foundation models and specifically optimized for high-quality multilingual translation tasks (2601.09012). The presented methodology leverages a two-phase approach: a supervised fine-tuning (SFT) stage using a diverse and large-scale parallel corpus, followed by reinforcement learning (RL) employing a composite reward signal from multiple learned and reference-based evaluation metrics. This report details the data curation strategy, the architectural and training modifications, the integrated RL reward composition mechanism, and robust empirical validation across extensive language coverage, including strong results for both text and image-situated translation.

Data Curation and Mixture Design

The construction of the SFT dataset emphasizes broad language coverage and high translation fidelity through a blend of human-curated and synthetic parallel data. Synthetic data generation employs Gemini 2.5 Flash for producing high-quality translations from the MADLAD-400 monolingual corpus, using quality estimation (QE) reranking and score-based selection to maximize downstream translation benefit and linguistic diversity. Sentence and longer text blob samples are included to enable both segment-level and document-level translation capabilities.

Human-authored parallel corpora from SMOL and GATITOS datasets supplement the mixture and boost coverage for lower-resource languages, achieving enhanced robustness for a spectrum of scripts and typologies. The data mixture for both SFT and RL is illustrated in the following figure, displaying the proportional language distribution as model tokens:

Figure 1: The SFT data mixture distribution, quantifying the token allocation among language groups used during training.

A significant fraction (30%) of the SFT mixture includes generic instruction-following data derived from the base Gemma 3 corpus, intended to regularize the model against catastrophic forgetting and preserve general-purpose reasoning abilities.

Supervised Fine-Tuning Procedure

The SFT phase employs Kauldron tooling with the AdaFactor optimizer, targeting all Gemma 3 checkpoints (4B, 12B, and 27B). All model weights are updated with the exception of embeddings, which are frozen to enhance generalization, particularly for scripts not seen during SFT. The SFT phase proceeds for 200k steps and batch size 64 at a fixed learning rate of 1e-4. The prompt used for translation inference is specialized to elicit highly accurate and idiomatic translations with explicit domain and structure conditioning.

Reinforcement Learning with Composite Metric Rewards

The RL enhancement modifies the SFT-pretrained checkpoints using a sophisticated reward composition. Several learned translation quality metrics are employed in an ensemble as reward models, including:

• MetricX-24-XXL-QE, a regression-based quality estimator producing fine-grained, scalar outputs.
• Gemma-AutoMQM-QE, an AutoMQM model finetuned from Gemma 3-27B-IT on MQM ratings, contributing span-wise reward signals.
• Lexical overlap metric ChrF, rescaled to align with the MQM-derived reward magnitude.
• A naturalness evaluation Autorater (LLM-as-a-Judge), penalizing outputs perceived as unnatural by native speakers.
• A generalist reward model for robustness against instruction-following and reasoning regressions.

A distinctive aspect is the additive integration of sequence-level and token-/span-level reward signals during RL advantage computation. This enables fine-grained error attribution and more granular credit assignment for both global and local translation phenomena.

Figure 2: Schematic illustrating additive combination of sequence-level and token-level rewards for RL advantage computation, applying uniform broadcast for sequence rewards.

Batch normalization is applied to combined advantages for stability during RL optimization.

Evaluation: Text and Image Translation

TranslateGemma undergoes extensive evaluation on automatic and human-centered MT metrics. Automatic metrics include MetricX24 and Comet22, and broad generalization is validated on the -.05em benchmark covering 55 language pairs. The results demonstrate unequivocal improvement for all model sizes compared to baseline Gemma 3. For instance:

• 27B TranslateGemma reduces MetricX from 4.04 to 3.09 (23.5% improvement).
• 12B TranslateGemma achieves 3.60 versus 4.86 baseline (25.9% reduction).
• 4B TranslateGemma outperforms the 4B baseline by 23.6% on MetricX.

Comet22 scores similarly improve, and these gains are consistent across both language-majority and low-resource settings, e.g., English-German drops from 1.63 to 1.19, English-Icelandic from 8.31 to 5.69, and English-Swahili from 5.92 to 4.45. Importantly, TranslateGemma 12B surpasses the baseline Gemma 3 27B on several metrics, establishing an efficiency benefit and cost reduction for equivalent translation quality.

Multimodal Generalization: Image Translation

TranslateGemma preserves and in some cases enhances the multimodal translation capability inherent in Gemma 3, despite absence of multimodal data in the SFT/RL phases. On the Vistra benchmark—which assesses visually-grounded translation—the 27B and 12B variants show substantial MetricX24 improvements for image-situated translation tasks.

Figure 3: Sample image from the Vistra benchmark used for evaluating image-text translation.

For example, the 27B model's MetricX improves by nearly 0.5 and its Comet22 by 1.6, substantiating robustness and cross-domain transferability.

Human Evaluation

MQM-based human evaluation with professional annotators across 10 diverse language directions on WMT25 establishes that TranslateGemma 27B and 12B largely outperform baseline Gemma 3 27B, except for German (comparable) and Japanese-to-English (marginal regression due to named entity errors), in full agreement with automatic metrics. Notably, strong reductions in error rates are observed for low-resource pairs such as English-Marathi and Czech-Ukrainian, confirming the practical utility of the training regimen.

Implications and Future Trajectories

TranslateGemma sets a reproducible benchmark for open MT model release, demonstrating that composite-metric RL atop SFT maximizes translation quality and efficiency across a highly multilingual landscape. The methodological blueprint—rich mixture SFT, composite reward RL with both global and local signal integration, and explicit instruction-following regularization—yields models that are both high-performing and generally capable.

The value of integrating high-quality synthetic data and employing span-level credit assignment RL is reinforced, with implications for future multilingual NMT architectures. Further, the models' resilience in zero-shot or few-shot multimodal layouts (without multimodal SFT) points to advances in emergent capabilities driven by strong pretrained vision-text features.

Future research directions include scaling to larger parameter counts, generalization to cross-domain document-level translation, better handling of named entities in challenging language directions, targeted multilinguality for underrepresented scripts, and more direct multimodal-finetuning and evaluation.

Conclusion

TranslateGemma represents a systematically enhanced series of MT models derived from Gemma 3, documenting consistent and strong improvements due to rich SFT and carefully structured RL. These models offer an efficient performance/compute trade-off, robust broad-language coverage, and preservation of core general and multimodal abilities. The open release facilitates further inquiry and practical deployment in multilingual and multimodal MT research (2601.09012).