Skip to content

Tencent's New Open-Source Models Outperform Google Translate in Most Categories

Tencent's new models beat Google Translate in 30 out of 31 language combinations. They're efficient, support minority languages, and are now open source.

This picture contains a paper in which some text is printed in a different language. We even see...
This picture contains a paper in which some text is printed in a different language. We even see two men are standing in the picture. This picture might be taken from the textbook.

Tencent's New Open-Source Models Outperform Google Translate in Most Categories

Tencent has introduced two new open-source translation models, Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B, outperforming larger competitors like Google Translate and GPT-4.1 in most categories. The models, developed with a unique fusion approach and extensive training, offer real-time bidirectional conversation translation and personalized language learning.

The models, with 7 billion parameters each, require fewer computational resources and run on weaker hardware compared to larger foundational models. Yet, they deliver comparable or even better performance. Tencent's five-stage training process and a dataset of 1.3 trillion tokens for minority languages contributed to their success.

In an international comparison test, Tencent's models outperformed established services like Google Translate in 30 out of 31 tested language combinations. The models support bidirectional translation between 33 languages, including widely spoken and less frequently digitized languages. Notably, they focus on translation between Mandarin Chinese and ethnic minority languages in China, supporting Chinese, Kazakh, Uyghur, Mongolian, and Tibetan.

The Chimera model, using a fusion approach, combines multiple translation suggestions into an improved final translation, achieving an average improvement of 2.3 percent in standardized tests. However, the research institution behind this model remains unclear.

Tencent's Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B models, now open source on Hugging Face and GitHub, offer a significant advancement in translation technology. Their performance, efficiency, and focus on minority languages make them valuable tools in today's multilingual world.

Read also:

Latest