Teams needing local translation
Downloaded weights can keep inference inside a controlled environment when the deployment is secured and properly operated.
Independent tool overview
TranslateGemma at a glance
TranslateGemma is Google's family of downloadable translation models based on Gemma 3, offered in 4B, 12B, and 27B sizes for text and text-in-image translation across 55 evaluated languages.
Visit the official TranslateGemma site ↗
- Developer
- Google Translate / Google
- Model sizes
- 4B, 12B, and 27B
- Evaluated coverage
- 55 languages on WMT24++
- Inputs
- Text or images containing text
- Input context
- 2,000 tokens total
- Access
- Downloadable weights under Gemma terms
- Reviewed
- September 1, 2026
Overview
What TranslateGemma is
TranslateGemma is a model family for developers and researchers, not a polished translation website for consumers. Google publishes 4B, 12B, and 27B instruction-tuned weights that accept source and target language codes plus either text or an image, then generate translated text. The models can be downloaded from Hugging Face or Kaggle or deployed through compatible local, cloud, and Google Cloud tooling.
The core release covers 55 languages evaluated on WMT24++ and retains Gemma 3's vision input for translating text found in images. Google also trained on nearly 500 additional language pairs, but explicitly says it did not yet have confirmed evaluation metrics for that extended set at launch. The model card documents a 2,000-token total input context, so this is not a drop-in system for translating arbitrarily long documents.
TranslateGemma's efficiency results are promising, but benchmark averages do not certify any particular language pair, domain, dialect, or document. Production teams still need bilingual evaluation, terminology controls, privacy architecture, fallback behavior, and human review—especially for legal, medical, financial, safety, immigration, or other consequential content.
Use cases
Who TranslateGemma is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Translation researchers
Three sizes, published evaluation data, and a technical report make the family useful for benchmarking, adaptation, and language-pair research.
Edge and laptop prototypes
Google positions the 4B model for mobile or edge work and the 12B model for consumer laptops, subject to runtime, precision, and hardware constraints.
Text-in-image workflows
The multimodal interface can extract and translate visible text from an image for constrained use cases.
Organizations with specialized corpora
The models can be evaluated or adapted for a defined language pair and terminology set under the applicable Gemma terms.
Capabilities
Core TranslateGemma features
Three deployment sizes
Provides 4B, 12B, and 27B variants so developers can trade memory, throughput, latency, and translation quality.
55-language evaluated core
Google reports WMT24++ results across 55 languages spanning high-, medium-, and low-resource families.
Text translation
Accepts a text string with explicit source and target language codes and returns translated text.
Image translation
Accepts an image and returns translated text using the retained Gemma 3 multimodal capability.
Regional language codes
The official chat template supports two-letter language codes and regional variants such as en-GB.
Downloadable weights
Official checkpoints are distributed through Hugging Face and Kaggle after acceptance of the Gemma usage terms.
Multiple serving paths
The official model page documents Transformers and compatible serving options, while Google offers deployment through Vertex AI Model Garden.
Published evaluation
The model card reports MetricX, COMET, MQM, and image-translation results for specified datasets and variants.
Process
How the TranslateGemma workflow works
Step 1
Choose a language pair and risk level
Define source and target locales, content domain, expected volume, and the consequence of an incorrect translation.
Step 2
Accept and review the terms
Read the current Gemma Terms and Prohibited Use Policy, including redistribution requirements, before downloading or serving the model.
Step 3
Select a model size
Benchmark 4B, 12B, and 27B on target hardware rather than assuming Google's broad deployment labels guarantee acceptable latency or quality.
Step 4
Use the required template
Pass exactly one text or image item with supported source and target language codes through the official chat template.
Step 5
Evaluate real material
Build a bilingual test set containing terminology, names, numbers, formatting, ambiguity, dialect, sensitive content, and adversarial inputs from the actual use case.
Step 6
Add production controls
Implement chunking, glossary or translation-memory logic, confidence or review routing, observability, access control, and safe failure behavior.
Step 7
Keep humans in the loop
Have qualified reviewers approve consequential translations and continuously sample live output for regressions across each supported locale.
Cost
TranslateGemma pricing and free plan
Google distributes TranslateGemma model weights under the Gemma Terms rather than selling a TranslateGemma subscription. Access through Hugging Face requires accepting the license. The real cost is hardware, cloud inference, storage, engineering, monitoring, and human quality review; managed Vertex AI charges depend on the infrastructure and deployment configuration.
TranslateGemma 4B
Weights available under Gemma terms
Smallest variant, positioned by Google for mobile and edge deployment.
- Lowest resource requirement of the three official variants
- Text and image input
- 2,000-token total input context
- Runtime, quantization, hardware, and operational costs are separate
TranslateGemma 12B
Weights available under Gemma terms
Middle variant, positioned for consumer laptops and local development.
- Google reports it outperforming the Gemma 3 27B baseline on WMT24++ MetricX
- More compute and memory than 4B
- Text and image input
- Infrastructure and review costs are separate
TranslateGemma 27B
Weights available under Gemma terms
Largest official variant for maximum fidelity in the release family.
- Google says it can run on one H100 GPU or TPU in the cloud
- Highest reported benchmark quality of the three variants
- Text and image input
- Requires materially more compute and memory
Managed deployment
Usage and infrastructure charges vary
Serve through Vertex AI Model Garden or another compatible cloud stack.
- No single flat TranslateGemma API price was published on the model page
- Costs depend on selected hardware, uptime, scaling, region, storage, and networking
- Managed hosting does not remove application-level evaluation, privacy, or review obligations
Pricing checked . Check current pricing at the source ↗
Assessment
TranslateGemma strengths and limitations
Where it stands out
- Purpose-built translation tuning is more focused than prompting a general chat model
- Three sizes support practical quality, latency, and cost tradeoffs
- Official weights allow local or controlled-environment inference
- Published model card and technical report provide architecture, training, and benchmark context
- Core evaluation spans 55 languages rather than only a few high-resource pairs
- Image input enables text extraction and translation in a single model path
- The 12B result demonstrates strong efficiency relative to the larger Gemma 3 baseline
- Developers can adapt the system and own the surrounding evaluation and serving stack
What to consider
- TranslateGemma is a model, not a finished consumer app or fully managed translation workflow
- The total input context is only 2,000 tokens, requiring careful segmentation for long documents
- Only 55 languages were part of the core reported WMT24++ evaluation
- Google says nearly 500 additional language pairs were used in training but did not have confirmed evaluation metrics for that extended set at launch
- Average benchmark scores can hide weak performance for a particular direction, dialect, domain, or content type
- The image benchmark was filtered to images containing a single text, so results should not be generalized to complex layouts, handwriting, dense documents, or multiple regions
- Generated translations can omit, add, soften, intensify, or mistranslate meaning while remaining fluent
- Names, numbers, negation, units, legal terms, medical language, safety instructions, and culturally specific expressions need special testing
- The model does not provide a reliable, calibrated guarantee that a translation is correct
- Local deployment transfers responsibility for security patches, access control, model provenance, logging, capacity, and incident response to the operator
- Weights are subject to the Gemma Terms and Prohibited Use Policy, not an unrestricted public-domain release
- Production cost includes compute, engineering, observability, terminology management, evaluation, and qualified human review
- Privacy depends on the chosen deployment: local inference can reduce external data transfer, while a hosted stack introduces its own processors and retention terms
- Consequential translations should not be released without a qualified human reviewer
Compare
TranslateGemma alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Tiny Aya
Choose Tiny Aya when a compact multilingual generative model with broader advertised language coverage is more important than translation-specific tuning.
Explore Tiny Aya →Consumer
Translate with ChatGPT
Choose Translate with ChatGPT when end users need a ready-made text, voice, and image translation interface rather than self-hosted model weights.
Explore Translate with ChatGPT →Miscellaneous
Gemini 3.5 Live Translate
Choose Gemini Live Translate when the primary requirement is real-time spoken translation instead of batch text or image inference.
Explore Gemini 3.5 Live Translate →Questions
TranslateGemma FAQs
What is TranslateGemma?
TranslateGemma is a family of Google translation models based on Gemma 3. Developers can download 4B, 12B, and 27B variants to translate text or visible text in images.
Is TranslateGemma a Google Translate replacement?
Not directly. Google Translate is a finished consumer service. TranslateGemma is a set of model weights and documentation for developers who will build, host, evaluate, and operate their own translation workflow.
How many languages does TranslateGemma support?
Google reports core training and evaluation across 55 languages. It also trained on nearly 500 additional language pairs but said confirmed evaluation metrics were not yet available for that extended set at launch.
Which TranslateGemma size should I use?
Start with 4B for constrained hardware, 12B for a balance of quality and cost, and 27B for the highest reported fidelity. Benchmark every candidate on your own language pair, domain, hardware, latency target, and review process.
Can TranslateGemma run locally?
Yes. Google positions 4B for mobile or edge use and 12B for consumer laptops. Actual feasibility depends on runtime, precision or quantization, available memory, accelerator support, and performance requirements.
Can it translate images?
Yes. The models accept an image and return translated text. The published image benchmark used a constrained set with a single text region, so complex documents and layouts require separate testing.
Is TranslateGemma free?
The official weights are downloadable after accepting the Gemma terms; there is no separate TranslateGemma subscription listed. Running the model still incurs hardware, cloud, engineering, monitoring, and human-review costs.
Is TranslateGemma open source?
Google calls it an open model and publishes weights, but use and redistribution are governed by the Gemma Terms and Prohibited Use Policy. Teams should review those terms rather than assuming a standard unrestricted open-source software license.
How long can the input be?
The official model card states a total input context of 2,000 tokens. Longer documents need structure-aware segmentation and a process for preserving terminology, references, formatting, and cross-segment context.
Is TranslateGemma accurate enough for legal or medical translation?
Do not rely on benchmark averages for consequential material. Use a qualified professional translator or reviewer, preserve the source, test the exact language direction and domain, and prohibit automatic release when an error could affect rights, care, money, or safety.
Does local TranslateGemma keep data private?
Local inference can keep source content off an external model API, but privacy still depends on the complete system: logs, telemetry, storage, backups, access controls, updates, and any surrounding services.
What should a production evaluation include?
Use native bilingual reviewers and representative material covering terminology, ambiguity, names, numbers, negation, units, dialect, formatting, offensive content, injection attempts, and known failure cases. Measure each language direction separately.
Bottom line
Our TranslateGemma verdict
TranslateGemma is a compelling building block for teams that need downloadable, translation-focused models and are willing to operate the full quality and serving stack. The 12B model's efficiency and the availability of 4B and 27B variants make experimentation practical. The responsible path is narrow and empirical: choose one language direction, benchmark real content, add terminology and review controls, and expand only after native speakers verify performance.
Visit TranslateGemma website ↗Similar Consumer Tools
View all Consumer AI tools →
Slackbot
Slackbot - Slack's personal AI agent for work
Step3-VL-10B
Step3-VL-10B - StepFun's open-source SOTA vision language model
Gmail
Gmail - Google's email inbox, now infused with Gemini for AI-powered insights, actions, and improvements
ERNIE 5.0
ERNIE 5.0 - Baidu's omni-modal, top-ranked Chinese model
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.