Chinese AI models roll out World Cup prediction features
As the 2026 FIFA World Cup brings together more nations, more matches and more excitement, a competition has intensified off the pitch among China's leading artificial intelligence models, which are being put to the test in predicting the results of one of the world's biggest sporting events.
The 23rd edition of the World Cup, featuring 48 teams, is being hosted by the United States, Canada and Mexico. It opened on Thursday and runs through July 19.
Several Chinese large language models, including Qwen, DeepSeek, Kimi and MiniMax, have rolled out prediction features, turning the tournament into a new testing ground for AI-powered reasoning and data analysis.
"As one of the most-watched sporting events across the globe, the World Cup offers AI companies a rare opportunity to showcase the computing power and analytical skills of their LLMs to a wider audience," said Guo Tao, a member of the Chinese Association for Artificial Intelligence and a senior expert in AI.
Several AI platforms have come up with interactive campaigns. For instance, Moonshot AI's Kimi has launched a 1 trillion-token reward pool, allowing users to share prizes by correctly predicting match winners and the final champion. A token refers to the smallest unit of data processed by AI models.
Alibaba Group's Qwen has introduced a dedicated match prediction assistant, while also offering human-versus-AI prediction challenges.
However, the World Cup has also exposed the limitations of current AI models when it comes to analyzing and predicting the results of sporting events. For example, before the Group C opener between Brazil and Morocco on Sunday, major LLMs made predictions in favor of Brazil based on both historical data and statistical indicators. The match ended in a 1-1 draw.
Guo said that while AI can analyze historical data and statistical models, it still struggles to accurately predict real-world results, especially in sports.
He pointed out that soccer matches are influenced by a wide range of factors in the physical world, and such variables are highly uncertain and difficult to quantify using fixed AI models, making precise predictions inherently challenging.
The limitations of current AI models were also highlighted by Wang Zhongyuan, president of the Beijing Academy of Artificial Intelligence, at this year's BAAI Conference held last week.
Wang said that while LLMs have become increasingly capable of solving problems in the digital world, many challenges in the physical world remain beyond their reach. As a result, the next stage of AI development will gradually shift from "predicting the next token" to "predicting the next physical state", he added.
Asked why tech companies are rolling out AI prediction features for sports when the accuracy rate is relatively low, Guo, the expert from the Chinese Association for Artificial Intelligence, said the trend partly reflects the growing pressure of competition across the industry.
"As competition in the LLM market intensifies, technological differentiation is becoming increasingly difficult. Companies are eagerly seeking new channels to distinguish themselves from their rivals," he said.
As the AI technology matures, simply competing on the size parameter is not enough, Guo said. "The market is paying less attention to how large a model is and more attention to whether it can deliver valuable services in real-world scenarios and solve practical problems for users," he added.
Hu Yanping, a professor at Shanghai University of Finance and Economics, said that LLMs and AI agents are already evolving from conversation-oriented systems into task-oriented systems, while moving beyond pretraining toward continuous learning and broader real-world perception.
"Exploratory projects, such as World Cup match predictions, can help accelerate this evolution," Hu said. "A capability framework built around perception, interaction, decision-making and collaboration is what future task-oriented AI agents need."
lijiaying@chinadaily.com.cn
