Translate Gemma.
Credit: Google
Just a day ago, OpenAI introduced the Translate GPT model translate 46 languages, including Kannada, Hindi, Marathi, Tamil and Telugu, faster and more accurately than ever before.
Now, arch-rival Google unveiled the company's most advanced TranslateGemma model. It can translate 55 global languages.
To support different classes of devices, it is available in 4B, 12B, and 27B parameter sizes. It can be implemented in application programming interfaces (API)s for mobile apps, software programs, websites and more.
The 4B parameter-based model can be used on mobile devices, whereas the new 12B parameter-based TranslateGemma model is designed to run smoothly on personal computers, bringing research-grade power to laptops.
The 27B model is said to be built for maximum fidelity, capable of running on a single H100 GPU or TPU in the cloud.
The search engine giant says that TranslateGemma can considerably reduce the error rate compared to the baseline Gemma model in all languages, achieving improved quality with greater efficiency.
It should be noted that the new TranslateGemma is a multi-modal language model. It promises accurate results not just in terms of identifying languages in text, but also in audio and image samples as well.
Out of 55 languages, Google's new TranslateGemma model supports 11 Indian languages, including Kannada, Assamese, Bengali, Gujarati, Hindi, Malayalam, Marathi and more.
The TranslateGemma model is available for registered researchers and developers and can be tested in Kaggle website (here), which is touted to be the world's largest online community and platform for data science and machine learning.
Get the latest news on new launches, gadget reviews, apps, cybersecurity, and more on personal technology only on DH Tech.