Chinese artificial intelligence firm iFlytek on Thursday unveiled the latest version of its AI model, Spark 4.0, which the company benchmarked against OpenAI’s GPT-4 Turbo, saying it shows better performance than the US company’s most-advanced model in five aspects, despite lagging in terms of coding and multimodal abilities.
Why it matters: Trained exclusively on a home-grown computing platform that is co-built with telecom giant Huawei, Spark 4.0 can be seen as sending a signal that China’s AI industry has entered a new phase of self-sufficiency despite the US’s restrictions on the country’s access to the most advanced chips, according to iFlytek.
Details: The Hefei-based company finally fulfilled its commitment to upgrade its Spark model to the level of GPT-4 in the first half of 2024, eight months after the promise was made. At Thursday’s event, however, Chairman Liu Qingfeng acknowledged that the gap would “widen again”, possibly to more than a year, once GPT-5 is released, because of its larger parameter size, longer training time, and the increase in the amount of data used.
- “It is necessary to recognize the comprehensive gap between China and the US in the LLM in a scientific and rational way,” said Liu, “but also to have the confidence to catch up quickly to ensure there is no generation gap.”
- The company has seen over 400,000 overseas developers build products based on the Spark V3.5 model that was introduced in January, with 7.02 million developers having used it in total.
- In a key upgrade, the Spark app and its desktop version, both of which are by iFlytek’s model series, added a feature called “personal space” where foundation models are dedicated to generating individualized content that matches one’s personality. The mechanism operates on the premise that users can input their information on work, study, and life, or accumulate sufficient dialog records with AI bots for enhanced learning.
Context: While iFlytek said its Spark 4.0 model ranked number one in eight international mainstream test sets, rival Alibaba on the same day was also certified by the world-renowned open-source platform Hugging Face as topping the world’s open-source big model list for its Qwen-2, with Meta’s AI model ranked behind it in second.