Tencent’s AI Bot, Hunyuan, Outperforms ChatGPT with Superior Chinese Language Capabilities

Sep 8, 2023

Tencent, China’s leading internet company, has introduced its advanced artificial intelligence model, Hunyuan, to the business sector. This AI bot’s most significant feature is its superior command of the Chinese language, making it a powerful tool in a country where ChatGPT, although capable of understanding Mandarin, is not accessible.

Over 1.4 billion people globally speak Chinese as their first language, with the vast majority residing in China. Consequently, this presents a substantial user base for domestic enterprises like Tencent to cater to.

Tencent, also known for its all-in-one application, WeChat, argues that Hunyuan’s proficiency in Chinese surpasses not only other AI models but also human capabilities. As evidence, Tencent’s vice-president, Jiang Jie, pointed out that Hunyuan outperformed GPT-4 in the Chinese university entrance exam during a demonstration at the Shenzhen business summit.

In the Chinese Language Understanding Evaluation benchmark, Hunyuan scored an impressive 86.918, slightly above Alibaba’s AliceMind which scored 86.685, with human competitors trailing behind.

An additional advantage of Hunyuan, according to Tencent, is its reduced rate of errors or “hallucinations” – instances where AI models generate and present incorrect information as facts. Jiang claimed that their method reduces the hallucination rate by 30 to 50 percent compared to current market options, although this assertion could not be independently confirmed by Quartz.

ChatGPT, on the other hand, has been criticized for its challenges with Chinese punctuation and terminology, along with a tendency to spread more misinformation in Chinese than in English.

Tencent’s AI ambitions are reflected in the numbers. As of July this year, Hunyuan’s model contained 100 billion parameters, units of language and their connecting variables, compared to OpenAI’s GPT-3 AI model which contained 175 billion parameters in 2020 and Meta Platform Inc’s Llama 2 model with 70 billion parameters in 2023. Furthermore, Hunyuan’s pre-training data consisted of 2 trillion tokens, dwarfing GPT-3’s 300 billion in 2020.

Tencent has already implemented over 50 large language model-enabled industrial solutions in various sectors like finance, media, travel, and education, demonstrating its commitment to AI development.

However, Tencent is not the only Chinese tech giant making strides in AI technology. Baidu, SenseTime Group, Zhipu AI, Alibaba, and JD.com have all developed their own versions of AI models. Despite these advancements, potential hurdles such as US sanctions could impede their progress by preventing Chinese companies from purchasing AI-grade chips and chip-making tools from overseas.