The 2023 iFlytek Global 1024 Developer Festival will be held in Hefei, Anhui from the 24th to the 29th. At the main forum of today’s opening ceremony,iFlytek Spark Cognitive Large Model V3.0 has been officially released. All seven capabilities have been continuously improved. Compared with ChatGPT, it is said to surpass Chinese and equivalent to English.Specifically, in terms of seven major abilities, text generation increased by 7%, language understanding increased by 6%, knowledge question and answer increased by 7%, logical reasoning increased by 8%, mathematical ability increased by 3%, coding ability increased by 9%, and multi-modal ability increased by 7%.
Liu Qingfeng revealed,In terms of mathematical capabilities, iFlytek Spark 3.0 is far ahead of ChatGPT. In the fields of design solution generation, code completion, testing and error detection, Spark 3.0's capabilities have surpassed ChatGPT.
The on-site PPT showed the Hengping report of the National Economic Research Institute of the Development Research Center of the State Council:At present, the comprehensive capabilities of Spark's large model have surpassed ChatGPT, leading the country, and its performance is particularly outstanding in the education, medical, and legal industries.
According to the English task test set released by OpenAI on the official website, in the test from October 16 to 20, the English ability of Spark V3.0 was equivalent to the ChatGPT 48 task results. The two scores were 85.1% (Spark V3.0) and 84.3% respectively.
However, Liu Qingfeng also said that compared with GPT4, iFlytek "still has a learning process."
Liu Qingfeng pointed out that large cognitive models are opening a new era of growth, and large models are increasingly relevant to everyone.
Since iFlytek's Spark was officially opened to the public on September 5, Spark's various application users have now exceeded 12 million.
The AI development tool iFlyCode has also been iterated to version 2.0, which can achieve efficiency improvements of 50%, 37%, and 44% in the design phase, coding phase, and testing phase respectively.
According to Liu Qingfeng, with the help of iFlytek's iFlyCode tool, the development work that originally took programmers three months to complete platform migration can now be completed in just one month.
It is worth mentioning thatHuawei has also dispatched employees to the iFlytek headquarters to form a "Joint Special Forces Team" that will double the training and inference efficiency of the Spark large model within half a year.
At the scene, iFlytek and Huawei's "Flying Star One" platform based on the Ascend ecosystem was officially released.