【资料图】

By CUI Peng

Tencent unveiled its large language model "Hunyuan" Thursday at the company’s annual summit.

Dowson Tong, CEO of cloud and smart industries, said Hunyuan had over 100 billion parameters and was trained with over 2 trillion tokens, two metrics often used to measure AI models" power.

Tong said Hunyuan possesses powerful capabilities in understanding Chinese, creativity and logical reasoning, and reliable task execution. Customers are able to access Hunyuan through APIs but also use it as a base model for industrial scenarios.

Tencent has collaborated with over 11,000 external partners in more than 100 sectors. Last year, the revenue generated by ecosystem partners accounted for one-third of Tencent Cloud"s overall revenue, with a growth rate four times that of Tencent Cloud"s overall growth.

JIANG Jie, vice president of Tencent, said Hunyuan is now connected to more than 50 of Tencent’s business sectors. He said that most large language models in the industry are still limited, to leisure scenarios with high fault tolerance in simple tasks. Jiang said Hunyuan has managed to reduce such “delusional answers” by at least 30 percent.

Hunyuan is capable of literary composition, text summarization and role-playing. It fully comprehends user intent and provides timely responses efficiently and accurately.

The latest version of Tencent Meeting based on Hunyuan can perform complex tasks such as extracting meeting information and analyzing content with simple natural language commands. It can also generate intelligent summaries minutes after the meeting.

In document processing, Hunyuan supports dozens of text creation scenarios. It can generate standard format text with one click, is proficient in over a hundred Excel formulas, supports natural language function generation, and can generate charts based on table contents.

In the advertising business context, Hunyuan supports intelligent ad material creation. It can combine industry and regional characteristics to achieve integration of text, images, and videos.