In the annual selection of the "Top Ten National Treasures" initiated by the State-owned Assets Supervision and Administration Commission of the State Council, China Telecom's independently developed Xingchen large model has successfully made the list due to its groundbreaking technological achievements. As the country's first full-scale, multi-modal, and entirely domestically produced foundational large model system, the Xingchen large model demonstrates exceptional capabilities in semantics, speech, vision, and multi-modal areas.
In the field of semantics, the Xingchen large model has achieved significant breakthroughs. Leveraging a fully domestic WanKa cluster and training framework, this model achieves over 93% of the computational efficiency of Nvidia's computing power, with training time accounting for as much as 98%. Notably, this model ranks first in the international authoritative evaluation list C-Eval and has surpassed several well-known domestic large models on the OpenCompass list.
In terms of speech technology, the Xingchen large model showcases unique advantages. As the first voice large model in the country to obtain dual filing from the Cyberspace Administration, it has successfully overcome the technical challenges of mixed dialect recognition, covering 40 dialect regions. Its supernatural voice generation technology achieves over 90% voice similarity with just 5-10 seconds of sample, reaching an industry-leading level.
In the visual field, the Xingchen large model has received an evaluation from academic experts stating that it "reaches an internationally advanced level." Currently, this model has been deployed on 765,000 cameras nationwide, with an average daily usage exceeding 600 million times, playing a significant role in urban governance, traffic management, and other areas.
In terms of multi-modal capabilities, the Xingchen large model also performs remarkably. Its image generation capability has surpassed the internationally renowned platform Midjourney, and its video generation technology ranks first on the VBench list. Notably, in long video generation, it has achieved coherent generation of 5-minute videos ahead of others.
Currently, the Xingchen large model has been transformed into over ten standardized AI products, applied in 50 industry scenarios including government affairs and industry. As a fully autonomous and controllable domestic AI system, the success of the Xingchen large model not only showcases China Telecom's technological strength but also marks significant progress in our country's AI field regarding core technological breakthroughs.
