TranslateGemma (blog.google/innovation-and-ai/technology/developers-tools/translategemma/) is a family of open-weight translation models developed by Google. Built directly on the Gemma 3 architecture and available in 4B, 12B, and 27B parameter variants, TranslateGemma delivers translation fidelity across 55 core languages (and nearly 500 experimental language pairs). By distilling knowledge and translation intuition from Gemini into compact open models, the 12B variant outperforms the baseline Gemma 3 27B model while retaining multimodal visual-text translation capabilities directly within images.

Deploying enterprise-grade machine translation at scale often forces organizations to choose between expensive black-box cloud APIs and underperforming local open-source models. TranslateGemma solves this tradeoff by providing high intelligence-per-parameter translation checkpoints that can be self-hosted privately on mobile edge chips, consumer workstations, or cloud GPU clusters.

Model Lineup & Architecture Specifications

Model Size	Target Hardware & Environment	Core Translation Strengths
TranslateGemma 4B	Mobile devices, Edge NPUs, lightweight CPU/GPU instances	Low memory footprint, fast edge latency, reliable for structured text.
TranslateGemma 12B	Mid-tier GPUs, workstation laptops, Apple Silicon	Sweet-spot balance: Outperforms the Gemma 3 27B baseline using $<50\%$ of parameters.
TranslateGemma 27B	Multi-GPU servers, high-throughput cloud clusters	Maximum nuanced context handling, idiomatic fidelity, and complex document translation.

Key Capabilities and Technical Innovations

• Two-Stage Gemini Knowledge Distillation & RL: Trained via a specialized two-stage pipeline combining Gemini synthetic data with reinforcement learning guided by translation quality reward ensembles (including MetricX-QE and AutoMQM) to reduce translation error rates significantly.

• Multimodal Visual Text Translation: Retains the vision-language foundations of Gemma 3. TranslateGemma extracts and translates text embedded inside images (e.g., signs, infographics, scanned documents) without requiring separate OCR preprocessing pipelines.

• Broad Coverage Across 55+ Language Pairs: Rigorously benchmarked across 55 major and low-resource languages (e.g., Spanish, French, Chinese, Hindi), with extended training spanning nearly 500 language pairs to serve as a base for downstream community fine-tuning.

• Open-Weight Ecosystem Availability: Weights are openly downloadable via Kaggle, Hugging Face, Ollama, and Google AI Studio / Vertex AI Model Garden for local private inference, on-premise hosting, and custom domain adaptation.