Google rolled out the TranslateGemma model family, and this is probably the most exciting release of the week for those who like to run neural networks locally.
The main feature here isn't the translation itself, but the size. The lineup includes a 4 billion parameter (4B) model. For context: this is compact enough to fit in a modern smartphone's memory and run completely autonomously.
No servers, no sending data to the cloud, fully working in airplane mode.
At the same time, Google engineers claim that thanks to the Gemma 3 architecture, this little 4B model performs like 12B models used to. Translation quality remains at a very high level (supporting 55 languages). For those with beefier hardware, there are 12B and 27B versions.
The weights are already on HuggingFace, so we can expect them to appear soon in MLX for Macs and in various local apps on Android/iOS.
TranslateGemma weights: http://huggingface.co/collections/google/translategemma
Official technology presentation (short vertical format):
https://www.youtube.com/shorts/FEta5fU4wW8
---
And for the record: OpenAI on the exact same day (coincidence? I think not) rolled out their own OpenAI Translate service.
It's simply a web interface that works, obviously, only with internet. The only cool feature is Tone of Voice adjustment — you can ask it to translate text in a casual way or conversely, in a formal business style. But it's still cloud-based.
So in the battle of "convenient web service" vs "freedom and local inference" — I'm definitely on Google's side.
OpenAI service: https://chatgpt.com/translate
UPDATE: Thanks to Grisha, chief editor of Techno and factcheck Jesuit. He corrects me:
"They quietly launched (OpenAI Translate) last year, there just wasn't an official announcement until now.
Journalists just noticed it now.
Proof: https://web.archive.org/web/20251119103023/https://chatgpt.com/translate/"