Liu Qingfeng: iFLYTEK will be the national team for full-stack autonomous and controllable large models
Liu Qingfeng, Chairman of iFLYTEK, revealed at the company's annual meeting that in the face of an environment with limited computing power, high demands and fierce competition, iFLYTEK has, on the one hand, unswervingly promoted the upgrade of its autonomous and controllable general large model base, and on the other hand, promoted the implementation of large model applications through GBC linkage. iFLYTEK Spark currently has a complete set of solutions from top-level planning to implementation.
Liu Qingfeng clarified the strategic ecological niche of iFlytek Spark Big Model: adhere to independent control, meet the rigid needs of people's livelihood, meet national strategic needs, and enable scientific research and production. "Focus on independent and controllable base capabilities and key application tracks, and maintain strategic focus with a long-term spirit." He revealed that in the new stage of general artificial intelligence development, iFlytek will become a national team for full-stack independent and controllable big models.
Previously, after the release of iFlytek Spark 4.0 Turbo, it surpassed GPT-4o in mathematical and coding capabilities. Released by authoritative domestic and foreign organizations such as the Allen Institute for Artificial Intelligence and OpenAI, iFlytek Spark 4.0-Turbo surpassed GPT-4o in 9 of the 14 mainstream test sets covering different task types such as understanding and reasoning, comprehensive exams, mathematics and science, and coding, with a relative efficiency improvement of 50%.
According to the comparison of the results on the code generation HumanEval test set, iFlytek Spark 4.0 Turbo has a slight gap with GPT-4o in tasks such as Python, Java, and JavaScript, and exceeds GPT-4o in C++ capabilities. In real application scenarios, based on the iFLYCode-Eval-2.0 code practical scenario test set built by the National Key Laboratory of Cognitive Intelligence, iFlytek Spark 4.0 Turbo has surpassed GPT-4o in tasks such as code generation, code error detection, and unit testing. At the same time, the Spark Code 7B version was launched to meet the needs of extremely fast response tasks such as code generation and code completion, with the best results in the industry.
