TranslateGemma
Open translation on Google models, supporting 55 languages
334 followers
Open translation on Google models, supporting 55 languages
334 followers
TranslateGemma is a new suite of open AI translation models built on Google’s Gemma 3. It enables high-quality communication across 55 languages, combining strong accuracy with exceptional efficiency. Designed to run on mobile, local devices, and cloud environments without compromising performance.
Alconost Localization
Checked it out with multiple languages and then evaluated the quality with our linguists. Some of the languages performed well, but others had quite many errors. Still, we've only checked it with one type of content, need to try it with different types.
Btw, why does it offer Arabic dialects but no Arabic (MSA) option?
This is a really meaningful step for open translation.
Running high-quality models locally (especially on mobile) solves both cost and privacy concerns that a lot of teams struggle with today.
I’m curious how this performs for long-form content (docs, blog posts, UI strings) versus short sentences — and whether there are plans to support domain-specific tuning (like technical or legal language).
Great to see Google backing truly open, practical models like this. Nice work.
The 4B model for mobile inference is interesting.... definitely want to try it on a Pi. Would love to have local translation that doesn't phone home.
The ensemble reward model approach is interesting using multiple quality signals (MetricX, AutoMQM, ChrF, naturalness) rather than optimizing against a single metric. We've seen very big reliability gains using multi-model consensus at inference time.
Really excited about the open-source approach for translation. I'm building a tech news aggregator that summarizes articles from Japanese engineering blogs for global readers, so local translation models like this could be game-changing for keeping latency low. Curious about the quality benchmarks for technical/specialized content vs general text — does the ensemble reward model help with domain-specific terminology?
Been waiting for something like this. Most translation APIs either cost a fortune or need internet. Having a local model that runs on Gemini 3 and actually handles 55 languages without phoning home? That's huge for anyone building offline-first apps or dealing with privacy-sensitive translation work.
Quick question though how's the accuracy compared to Google Translate's API for less common language pairs? Like, does it handle nuanced stuff well or is it better for straightforward translation?
Either way, open source translation models are a game changer. Nice work 🔥