China releases Spark X1 deep reasoning model that packs a math computation punch

China on Wednesday released a deep reasoning large language model based on the current domestic computing power, Spark X1.

The model was jointly developed by the country's two AI giants, iFLYTEK and Huawei. Spark X1 ranks number one in the country in terms of Chinese mathematics computation capabilities and has been applied in education, medical scenarios, and other contexts.

At the press conference, researchers demonstrated Spark X1's performance in answering college entrance examination questions, American Invitational Mathematics Examination (AIME) competition questions and high school Olympic questions. Spark X1 not only accurately answered the questions but also detailed the ideas and steps for solving the problems. Insiders say that compared with general large models, its problem-solving process is closer to the "slow thinking" method of humans.

Spark X1 demonstrated the three typical characteristics of deep reasoning models: simplifying complex problems, conducting self-exploration and reflection verification, as well as strengthening training based on accurate-or-not feedback.

The researchers said Spark X1 even surpassed ChatGPT o1 in some "exams" it recently participated in, including primary, middle, and high school competitions, university mathematical tests, AIME, and MATH 500.

At present, Spark X1 has been applied in education. About a hundred pilot areas in China have reported that Spark X1 has performed well in providing multiple solutions to a single problem, associating teaching knowledge, and expanding students' critical thinking.

In the medical field, Spark X1 has achieved initial verification results, which can achieve an accuracy rate of 90 percent for specialist auxiliary diagnosis and complex medical record content quality control.

(Cover: Inside the Beijing iFLYTEK building. /CFP)