Google's Open-Source Translation Model
TranslateGemma is a suite of lightweight, open translation models built on the Gemma 3 foundation. Delivering high-quality machine translation across 550 languages with multi-modal capabilities, TranslateGemma enables developers and researchers to build powerful cross-language applications with state-of-the-art performance.
Purpose-built for high-quality, efficient machine translation
Reliably translate across 550 languages covering high, medium, and low-resource languages. Additionally trained on nearly 500 language pairs for expanded research capabilities.
Translate text within images without additional fine-tuning. Perfect for signs, menus, documents, and real-world visual content translation.
Choose from 4B (mobile/edge), 12B (laptop), or 27B (cloud) parameter models. The 12B model outperforms the 27B Gemma 3 baseline.
Fully open models available on Hugging Face, Kaggle, Vertex AI, and Ollama. Comprehensive documentation and code examples included.
TranslateGemma outperforms baseline models on WMT24++ benchmarks
| Model Size | Model | MetricX (lower is better) | Comet22 (higher is better) |
|---|---|---|---|
| 27B | Gemma 3 | 4.04 | 83.1 |
| TranslateGemma | 3.09 | 84.4 | |
| 12B | Gemma 3 | 4.86 | 81.6 |
| TranslateGemma | 3.60 | 83.5 | |
| 4B | Gemma 3 | 6.97 | 77.2 |
| TranslateGemma | 5.32 | 80.1 |
The 12B TranslateGemma model achieves better performance than the 27B Gemma 3 baseline, demonstrating significant efficiency gains through specialized fine-tuning. This means higher translation quality at lower computational cost.
Built on Gemma 3 with innovative two-stage training
TranslateGemma is built directly on the Gemma 3 foundation model, inheriting its powerful multilingual capabilities and efficient architecture. The model retains Gemma 3's multi-modal abilities while being specifically optimized for translation tasks.
Through specialized fine-tuning, TranslateGemma transforms general language understanding into precise cross-language conversion, achieving superior accuracy and fluency in translations.
Training on large-scale parallel data combining synthetic translations from Gemini 2.5 Flash and high-quality human translations.
Optimization using composite reward models including MetricX-QE and AutoMQM for human-aligned translation quality.
From mobile apps to enterprise solutions
Build real-time conversation translation apps, localize websites and documents, analyze multilingual market data.
Develop apps that instantly translate street signs, restaurant menus, product labels, and scanned documents.
Deploy privately for data security. Reduce operational costs with efficient models while maintaining quality.
Use as a foundation for automatic post-editing, domain-specific fine-tuning, and cross-language innovation.
550
Languages Supported
How TranslateGemma compares to Meta's NLLB (No Language Left Behind)
| Aspect | TranslateGemma (Google, 2026) | NLLB-200 (Meta, 2022) |
|---|---|---|
| Architecture | Built on Gemma 3 (decoder-only LLM), fine-tuned via SFT + RLHF-style reinforcement learning with MetricX-QE, AutoMQM reward models | Dedicated encoder-decoder Transformer; trained specifically as a many-to-many MT model |
| Model Sizes | 4B, 12B, 27B parameters | ~3.3B dense / 54.5B MoE; distilled variants ~600M-1.3B |
| Languages | 55 core languages (high-, mid-, low-resource); trained on ~500 pairs | 200+ languages (strong emphasis on low-resource/underrepresented) |
| Performance | 12B beats 27B Gemma 3 baseline (~25% error reduction); SOTA on WMT24++ for 55 languages | SOTA for 200-language coverage in 2022-2024; strong on rare pairs |
| Efficiency | Highly efficient; 12B outperforms 27B baseline; 4B suitable for mobile/edge | MoE version needs significant hardware; distilled versions for modest hardware |
| Multimodal | Yes - translates text from images without fine-tuning (Vistra benchmark) | Text-only; no native vision/audio support |
| Best For | High-quality translation on 55 languages, local/offline deployment, image+text translation | 200+ languages, rare/low-resource languages, many-to-many without English pivot |
You need top-tier quality + efficiency on ~55 languages, want to run locally on phones/laptops, or need image text translation. The 12B version delivers exceptional performance-to-cost ratio.
You need 200+ languages (particularly rare ones), many-to-many translation without routing through English, or working with legacy MT setups where NLLB is already integrated.
Common questions about TranslateGemma
ollama run translategemma). All models are released under the Gemma Terms of Use.
Multiple ways to access TranslateGemma
Model weights & transformers integration
Notebooks & experiments
Enterprise cloud deployment
Local deployment
# Quick start with Hugging Face Transformers
from transformers import pipeline
import torch
pipe = pipeline(
"image-text-to-text",
model="google/translategemma-12b-it",
device="cuda",
dtype=torch.bfloat16
)
# Or run locally with Ollama
# ollama run translategemma