iFLYTEK upgrades Spark large model and continuously promotes the industrial application of AI large models | Frontline

Written by | Wang Fangyu

Edited by | Su Jianxun

iFLYTEK, known as the "National Team of General Large Models", has once again upgraded the capabilities of its large model.

On October 24th, at the opening ceremony of the 7th World Sound Expo and the 2024 iFLYTEK Global 1024 Developer Festival, Liu Qingfeng, the chairman of iFLYTEK, released the XF Spark 4.0 Turbo.

Liu Qingfeng introduced that the XF Spark 4.0 Turbo has been newly upgraded. According to the back-to-back tests based on fresh and real data, its seven capabilities comprehensively surpass GPT-4 Turbo, and its mathematical and coding capabilities exceed GPT-4o. It has achieved the first place in 9 out of 14 mainstream tests in Chinese and English at home and abroad.

At the conference, iFLYTEK also demonstrated the progress of the large model in multimodal interaction capabilities. It is understood that on the basis of the original far-field high-noise, full-duplex, and multilingual and multi-dialect capabilities, this upgrade has added multimodal capabilities, including super anthropomorphic and personalized capabilities, achieving a multimodal interaction that integrates voice, video, and graphics.

In terms of computing power, iFLYTEK has always insisted on building an independent and controllable general large model base based on domestic computing power. Last October, iFLYTEK jointly launched the first domestic large-scale model computing power platform with 10,000 cards, "Feixing No.1", in collaboration with Huawei.

At this conference, the domestic super-large-scale intelligent computing platform "Feixing No.2", jointly created by iFLYTEK, Huawei, and Hefei Big Data Asset Operation Co., Ltd., was also officially launched. Liu Qingfeng stated that the launch of the upgraded "Feixing No.2" will bring about the continuous adaptation of new models and new algorithms and another leap in the scale of the intelligent computing cluster, leading the development of the domestic large model base and providing the world with a second choice.

Since this year, a number of large model companies have stopped the pre-training process. The pre-training model is the underlying core technology of large model companies, and stopping usually means leaving the game. By upgrading the Spark model and launching a new intelligent computing platform, iFLYTEK conveys the attitude and confidence of continuous pre-training.

Liu Cong, the dean of the iFLYTEK Research Institute, said in an interview with 36Kr that iFLYTEK's large language model will insist on iteration and continue to make layouts in aspects such as similar to o1 large model, multimodal interaction, and end-to-end voice.

Liu Cong frankly admitted that iFLYTEK's computing power scale is not as large as that of the leading companies; using the domestic computing power platform also requires a lot of additional effort for adaptation and other work. However, iFLYTEK insists on building and upgrading an independent and controllable general base large model based on domestic computing power, and has achieved considerable results in the situation where the domestic chips and computing power clusters have a certain gap compared with the international leading level.

It is reported that in the past year and more of practice, the training and inference performance of the Spark large model on "Feixing No.1" has been continuously optimized, and the performance of some test sets even exceeds the internationally leading GPT-4 Turbo.

In the industrial application of large models, iFLYTEK has been actively promoting and is committed to "solving the urgent needs of society with artificial intelligence technology". Public information shows that from January to September this year, iFLYTEK successfully won 38 projects, and the disclosed winning bid amount was 216 million yuan. Both the number of projects and the amount are ranked first in the industry.

"The exploration of future AI technology must be large-scale industrialized and must enter the real deep-water area in the scenarios. All those who play with 'concepts' will not have a great future," Liu Qingfeng said in his speech.

At the conference, based on the capabilities of the XF Spark base, iFLYTEK released the latest product applications for multiple industry scenarios such as education, healthcare, judiciary, government services, and enterprise office. It is reported that as of October 2024, iFLYTEK has jointly built more than 20 industry large models with various leading enterprises, covering more than 300 application scenarios.