China's Tencent debuts large language AI model, says open for enterprise use (updated)

SenseTime

07/09/2023 05:45

(Adds details, context and exec quotes from paragraph 3 and
below)
    By Josh Ye
       BEIJING, Sept 7 (Reuters) - Tencent Holdings  0700.HK 
said on Thursday companies could now use its large language
artificial intelligence (AI) model "Hunyuan" as it premiered the
much-awaited product amid a race by tech firms race to become
China's AI champion.
    The Chinese tech giant and owner of the WeChat social media
platform conducted a demonstration before a live audience at a
conference in Shenzhen, and said Hunyuan had become the
foundation of more than 50 of its products and services.
    "By July, there are more than 130 large language models in
China," Jiang Jie, Tencent's vice president, said. "A war of a
hundred models has begun."
    Hunyuan's debut comes after several Chinese tech firms
including Baidu Inc  9888.HK  and SenseTime Group  0200.HK 
recently unveiled their own AI models.
    Tencent, China's most valuable internet company, said
Hunyuan had more than 100 billion parameters and was trained
with more than 2 trillion tokens, two metrics often used to
measure AI models' power.
    OpenAI's GPT-3 AI model contained 175 billion parameters in
2020 and Meta Platform Inc  META.O 's Llama 2 model had 70
billion parameters in 2023. 
    Tencent said its model, capable of conversing in both
Chinese and English, was "better" than OpenAI's ChatGPT in areas
such as writing long text with thousands of words and solving
certain math problems. 
    The model also experiences 30% less hallucination compared
to Llama 2, Tencent added. AI experts often describe moments
where AI models generate incorrect information but present it as
if it was a fact as "hallucination".
    There was no independent verification of Tencent's claims.
ChatGPT and Meta were not immediately available for comment.


 (Reporting by Josh Ye in Hong Kong, Additional reporting by the
Beijing Newsroom; Editing by Christopher Cushing and Miral
Fahmy)
 ((qiaoyi.li@thomsonreuters.com;))