There is a lack of a "bridge" between National Core + National Model, and Qingcheng Jizhi has set up a road to leapfrog domestic computing power at WAIC for two consecutive years.

Smart Things Author | Wang Han Editor | Moying Token, like water, electricity and coal in the industrial era, is evolving into the core production factor and value carrier in the AI ​​era. However, every token generation depends on the underlying computing power support and the upper model core. If China's AI industry wants to take the initiative in the Token era, it must achieve autonomy and controllability in these two core links. At the national core level, domestic computing power chips such as Huawei Ascend and Haiguang have moved towards large-scale deployment, the computing base has begun to take shape, and the measured performance continues to approach the international mainstream level; at the national model level, domestic large models such as DeepSeek, GLM, Qwen are on par with or even surpass the international mainstream level in text generation, multi-modal understanding, and reasoning capabilities, and their intelligence level is becoming increasingly powerful. Both Guoxin and Guomo are in place, but there is a key link between them that has been ignored for a long time: the inference engine. This is a key link that determines the efficiency and quality of Token production. The mainstream method is still based on foreign frameworks. Qingcheng Jizhi has always been committed to the construction of this link. At last year's WAIC conference, Qingcheng Jizhi concretely demonstrated the role of its Chitu large model inference engine in lowering the threshold of inference hardware. This year, with the surge in demand for tokens and the birth of the new product AI Ping, Qingcheng Jiezhi officially released a complete domestic Token "front store and back factory" integrated product system, using the self-developed Chitu reasoning engine as the core base of the "back factory" production, the AI ​​Ping one-stop Token service platform as the "front store" circulation and dispatching hub, and the Bagua furnace series software as the terminal implementation tool. Qingcheng Jizhi wants to use practice to prove to the industry that the full link of "domestic chips + domestic reasoning framework + domestic large models" is not only feasible, but also lower cost, more efficient, and autonomous and controllable. Regarding the domestic Token "front store and back factory" model, the dilemma and breakthrough of domestic chips, etc., Zhidongxi held an in-depth dialogue with Dr. Shi Tianhui, co-founder of Qingcheng Jizhi, and Dr. Tang Shizhi, co-founder and head of Chitu Inference Engine. 1. Self-developed from the first line of code, Chitu is in China