Overview
TranslateGemma is a lightweight open-source translation model Google released on September 15, 2026, trained from Gemini, supporting 55 languages and able to run offline. It was announced alongside a batch of Google language technology results; Google says its language technology now covers more than 300 languages, reaching roughly 86% of the world's population.
Its role is to close a gap that large-model translation leaves open. General models translate well, but calls require a network connection, are billed by volume and involve sending data outside. A dedicated translation model can run on a local device, which suits documents that must stay confidential, unstable networks, or simply a wish to bring the cost of bulk translation down. The 55 languages cover the major ones, forming a division of labour with the 300-plus language reach that depends more on online services.
Within Google's open model line-up TranslateGemma sits under the Gemma family, sharing origins with general multimodal models such as Gemma 4 but aiming elsewhere: Gemma 4 handles general reasoning and multimodality, while TranslateGemma focuses on translation and tries to do that one job small and well.
Key Features
- Trained from Gemini: Distils Gemini's translation ability into a lightweight model, balancing size against quality.
- 55 languages: Covers the major languages, which is enough for common multilingual translation needs.
- Runs offline: Weights can be downloaded to a local device, so translation does not depend on a network and data stays in place.
- Open weights: The model is downloadable, so it can be self-hosted, reviewed and modified.
- Part of the Gemma family: Shares origins with Gemma 4, so existing local deployment tooling and experience carry over.
Use Cases
- Confidential document translation: content stays on the local device and never leaves the internal network
- Privacy-sensitive material: internal email, legal documents and medical records in multiple languages
- Translation needs in unstable-network or fully offline environments
- On-the-spot translation on phones, laptops and other edge devices
- Bulk translation where cost also has to be controlled
Pros
- Runs offline, which makes it usable for confidential and privacy-sensitive work
- Open weights leave deployment options open
- Trained from Gemini, so the starting point is Google's existing translation ability
- Shares origins with the Gemma family, so local deployment experience transfers directly
- Lightweight enough to run on edge devices
Pricing
Open source and free, with weights available for download. The running cost is local hardware usage, with no metered billing.
Summary
TranslateGemma is the lightweight open-source translation model Google released on September 15, 2026, trained from Gemini, supporting 55 languages and able to run offline. It pulls translation out of the networked large-model call and hands it to a dedicated small model that runs on a local device, which suits confidential documents, privacy-sensitive material and constrained networks. As a member of the Gemma family it shares origins with Gemma 4, so existing local deployment experience carries over directly.