Google has officially announced TranslateGemma, a new suite of open translation models designed to enable efficient multilingual translation across 55 languages. Built on the Gemma 3 architecture, the models are targeted at researchers and developers seeking open, locally deployable translation systems instead of closed, cloud-only solutions.
With this launch, Google is positioning TranslateGemma as a strong alternative to popular AI-powered translation tools, including ChatGPT’s translation capabilities, by emphasising openness, efficiency, and control over deployment.
TranslateGemma is a collection of translation-focused AI models derived from Google’s Gemma 3 family. The models are optimised specifically for multilingual translation tasks while maintaining a relatively compact footprint compared to larger general-purpose language models.
TranslateGemma is released in three parameter sizes:
4B parameters – suited for mobile and edge devices
12B parameters – designed for balanced local or server deployments
27B parameters – aimed at high-performance cloud or enterprise use cases
This range allows developers to choose models based on hardware constraints and translation complexity.
According to Google, TranslateGemma models are trained using a two-stage approach:
Supervised fine-tuning
Reinforcement learning
The models leverage a mix of:
High-quality human-generated translation data
Carefully curated synthetic translation data
Google states that TranslateGemma reduces translation error rates across:
High-resource languages
Mid-resource languages
Low-resource languages
This improvement is achieved while using fewer parameters compared to baseline Gemma models.
ChatGPT’s translation feature, powered by OpenAI’s models, is widely used for conversational and real-time translation but operates as a closed, cloud-based system.
TranslateGemma follows a fundamentally different approach.
TranslateGemma offers open model weights
Developers can download, inspect, fine-tune, and deploy the models
Models can run on:
Local devices
Private servers
Custom hardware environments
This design allows organisations working with:
Sensitive data
Regulated industries
Low-connectivity or offline environments
to deploy translation systems without sending data to external servers, directly contrasting with ChatGPT’s cloud-first workflow.
TranslateGemma currently supports:
55 evaluated language pairs
Training exposure to nearly 500 additional language pairs for future experimentation
Google notes that TranslateGemma retains multimodal capabilities from Gemma 3.
This enables:
Text translation within images
Document and visual content translation
Multimodal use cases without separate multimodal training
TranslateGemma models can be accessed through:
Kaggle
Hugging Face
Google Colab
Vertex AI
Google has also released a detailed technical report outlining:
Training methodology
Benchmark results
Evaluated languages
Model performance characteristics
With TranslateGemma, Google is reinforcing the role of open translation models as a viable alternative to proprietary AI translation platforms.
Greater control over deployment
Customisation through fine-tuning
Improved data privacy
Flexibility across hardware environments
TranslateGemma represents Google’s most direct move yet to challenge ChatGPT Translate by offering open, efficient, and locally deployable translation models. By combining multilingual coverage, reduced error rates, and multimodal capabilities, TranslateGemma gives developers and enterprises a powerful new option for building custom translation systems beyond cloud-only AI services.