iFlytek launches large language model SparkDesk

By October, SparkDesk will be a universal model benchmarked against ChatGPT, and performing better in Chinese.

Photo from CFP

Photo from CFP

By JIANG Jingling


Chinese tech company iFlytek has unveiled its entry into the homegrown AI large language model race, SparkDesk, with a live demonstration.

LIU Qingfeng, chairman of iFlytek, said that by October, SparkDesk will be considered a universal model benchmarked against ChatGPT, surpassing the current version in Chinese and reaching a comparable level in English.

Speaking your language

Liu said in an interview that iFlytek has advantages in long-text generation and math. A Jiemian News reporter witnessed SparkDesk writing high-quality academic papers, promotional material, stories and other content.

SparkDesk performs well in the Chinese language, interpreting colloquialisms and answering emotional questions with emotional intelligence.

When asked routine high-school math problems, the AI quickly solves routine calculations.

Lots to learn

In terms of education, after being equipped with SparkDesk, AI learning machines can correct writing and conduct live conversations just like human teachers, with in-depth, high-level corrections and analysis of structure and style.

Speech-to-text transcription can automatically generate meeting summaries, organize complex audio recordings into coherent text, and generate news.

In terms of long text generation, iFlytek’s large model is significantly ahead in China and surpasses ChatGPT in Chinese, but still has a certain gap in English, Liu said.

Long way to go

LIU Cong of the iFlytek Research Institute told Jiemian News that compared with Alibaba and Baidu, iFlytek excels in large model algorithms. Since the launch of the “iFlytek Super Brain” in 2014, the team has amassed a large amount of data in industries such as education and healthcare.

However, Liu was clear that there are still plenty of flaws in large model technology, such as difficulties in updates, confusing one thing with another in factual questions, and fabricating historical facts.