TranslateGemma

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.

Why TranslateGemma

Purpose-built for high-quality, efficient machine translation

🌍

550 Languages Supported

Reliably translate across 550 languages covering high, medium, and low-resource languages. Additionally trained on nearly 500 language pairs for expanded research capabilities.

🖼️

Multi-modal Translation

Translate text within images without additional fine-tuning. Perfect for signs, menus, documents, and real-world visual content translation.

Three Model Sizes

Choose from 4B (mobile/edge), 12B (laptop), or 27B (cloud) parameter models. The 12B model outperforms the 27B Gemma 3 baseline.

🔓

Open & Accessible

Fully open models available on Hugging Face, Kaggle, Vertex AI, and Ollama. Comprehensive documentation and code examples included.

TranslateGemma Benchmark Performance

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

Key Insight

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.

TranslateGemma Technical Architecture

Built on Gemma 3 with innovative two-stage training

Foundation

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.

1

Supervised Fine-Tuning (SFT)

Training on large-scale parallel data combining synthetic translations from Gemini 2.5 Flash and high-quality human translations.

2

Reinforcement Learning (RL)

Optimization using composite reward models including MetricX-QE and AutoMQM for human-aligned translation quality.

TranslateGemma Use Cases

From mobile apps to enterprise solutions

Cross-Language Communication

Build real-time conversation translation apps, localize websites and documents, analyze multilingual market data.

Visual Content Translation

Develop apps that instantly translate street signs, restaurant menus, product labels, and scanned documents.

Enterprise Solutions

Deploy privately for data security. Reduce operational costs with efficient models while maintaining quality.

Research & Development

Use as a foundation for automatic post-editing, domain-specific fine-tuning, and cross-language innovation.

550

Languages Supported

English Chinese (Simplified) Chinese (Traditional) Spanish French German Japanese Korean Portuguese Russian Italian Dutch Polish Turkish Vietnamese Thai Indonesian Hindi Bengali Arabic Hebrew Ukrainian Czech Romanian Hungarian Finnish Swedish Danish Norwegian Greek + 25 more

TranslateGemma vs NLLB-200

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

Choose NLLB When:

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.

Frequently Asked Questions

Common questions about TranslateGemma

TranslateGemma is Google's open-source machine translation model suite released in January 2026. Built on the Gemma 3 foundation model, it delivers high-quality translation across 55 languages with multimodal capabilities (text in images). Available in 4B, 12B, and 27B parameter sizes, it's optimized for efficiency—the 12B model outperforms the 27B Gemma 3 baseline while using fewer resources.
Gemma 3 is Google's lightweight, state-of-the-art open-weight model family released in early 2025, built on research from the Gemini series. It features multimodal capabilities (text + image), 128K context window, support for 140+ languages, and efficient architecture for single GPU/TPU deployment. TranslateGemma is built directly on Gemma 3, inheriting its multilingual and multimodal capabilities while being specifically fine-tuned for translation tasks through supervised learning and reinforcement learning.
Choose based on your deployment environment: 4B for mobile/edge devices and resource-constrained environments; 12B for consumer laptops and desktops—this is often the best balance as it outperforms the 27B Gemma 3 baseline; 27B for cloud servers and maximum translation quality when resources aren't limited. The 12B model is particularly recommended for most use cases due to its exceptional efficiency-to-performance ratio.
Yes, TranslateGemma inherits Gemma 3's multimodal capabilities and can translate text extracted from images without additional fine-tuning. This makes it ideal for real-world applications like translating street signs, restaurant menus, product labels, and scanned documents. The model performs strongly on the Vistra image translation benchmark even though it wasn't specifically trained for image translation.
TranslateGemma and Meta's NLLB-200 serve different needs. TranslateGemma offers superior quality and efficiency on its 55 supported languages with multimodal capabilities, ideal for local deployment. NLLB-200 supports 200+ languages with emphasis on rare/low-resource languages and many-to-many translation without English pivoting. Choose TranslateGemma for quality + efficiency on common languages; choose NLLB for maximum language coverage including rare languages.
TranslateGemma is available on multiple platforms: Hugging Face for model weights and transformers integration; Kaggle for notebooks and experiments; Vertex AI for enterprise cloud deployment on Google Cloud; and Ollama for easy local deployment with a simple command (ollama run translategemma). All models are released under the Gemma Terms of Use.
TranslateGemma uses an innovative two-stage training approach: First, Supervised Fine-Tuning (SFT) on large-scale parallel data combining high-quality synthetic translations from Gemini 2.5 Flash and human translations. Second, Reinforcement Learning (RL) optimization using composite reward models including MetricX-QE and AutoMQM to align translation quality with human preferences. This approach achieves ~25% error reduction compared to baseline models.
Yes, TranslateGemma is freely available as open-weight models released under the Gemma Terms of Use. You can download the model weights from Hugging Face or Kaggle, run them locally with Ollama, or deploy on cloud infrastructure. For commercial use, review the Gemma Terms of Use to ensure compliance with the license terms.

Get Started with TranslateGemma

Multiple ways to access TranslateGemma

# 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