According to preliminary results from the organizers, Hunyuan-MT-7B ranked first in 30 out of 31 tests, surpassing even famous systems from Google and OpenAI.
WMT25 is an annual event that brings together machine translation researchers and developers worldwide, considered the "top playground" to evaluate translation technology capabilities.
What caught the attention of the technology world was the "small but mighty" scale, with only 7 billion components of the Hunyuan-MT-7B.
Instead of following the trend of enlarging the number of parameters, Tencent focuses on building a “comprehensive training process”, including data selection, continuous training, multi-layer fine-tuning, and reinforcement learning.
As a result, the model still produces smooth, accurate results and even outperforms competitors that are many times larger.
"The strength of Hunyuan-MT-7B is that it uses fewer parameters to achieve quality equal to, or even better than, giant models," the development team shared.
The languages tested range from Arabic to Estonian to Maasai, spoken by about 1.5 million people in Kenya and Tanzania.
There are also Czech-Ukrainian and Japanese-Simplified Chinese language pairs. Only Bhojpuri – a language spoken by more than 50 million people in India and Nepal – did Hunyuan-MT-7B not reach the top spot.
However, the organizers also emphasized that these are only preliminary results, based on automatic assessment.
“These numbers need to be viewed with caution, as there may still be bias in the scoring process,” said Tom Kocmi, a researcher at Cohere and head of the WMT25 organizing committee.
The official results based on manual human evaluation will be announced at the conference taking place from November 5 to 9 in Suzhou city, China.
Not just stopping at the lab, Tencent has brought this model into real products such as the online meeting platform Tencent Meeting, web browser and enterprise version of the WeChat application.
At the same time, Hunyuan-MT-7B has also been shared on Hugging Face and GitHub for the AI development community to exploit.
This event once again shows that the AI machine translation race is heating up every day and Tencent's "dark horse" Hunyuan-MT-7B has proven that it is not the big model that is strong.
(According to aitechsuite)
Source: https://vietnamnet.vn/trung-quoc-bat-ngo-qua-mat-my-tai-cuoc-thi-dich-thuat-ai-toan-cau-2438801.html
Comment (0)