Kechang Xunfei's Large Model Training Speeds Up Again! Domestic Computing Power Has Reached A100 Level!

On the recent interactive platform, iFLYTEK announced an exciting development: after years of effort and investment, they have made significant progress in the training and inference efficiency of their Spark large model. Unlike traditional methods that directly rely on NVIDIA GPUs, iFLYTEK chose a more challenging domestically-produced computing power solution. Under limited resources, their team collaborated with Huawei to successfully overcome multiple technical challenges.

Since May 2023, the collaboration between iFLYTEK and Huawei has broken through several technical bottlenecks, including high-speed interconnection networking for thousands of cards, optimization of computing communication, and the realization of high-throughput inference. These efforts have resulted in the training efficiency of iFLYTEK's general large model and deep reasoning model increasing from 30%-50% to 85%-95%, performing excellently and almost comparable to NVIDIA's A100 GPU.

In 2025, iFLYTEK further embraced challenges and successfully solved two major problems in domestic computing power training. On one hand, they improved the training efficiency of long-term reasoning chain reinforcement learning, raising it from 30% compared to A800 to over 84%; on the other hand, they achieved breakthroughs in the full-chain training efficiency of MoE models, with training efficiency soaring from 30% in March of this year to 93%. These achievements not only mark a major advancement in domestic computing power technology but also laid a solid foundation for iFLYTEK in this field.

With the continuous improvement of domestic computing power technology, iFLYTEK's training costs are expected to drop significantly. This will bring more possibilities for the company's future development and set an example for China's technological industry's independent innovation.

In summary, these achievements by iFLYTEK not only demonstrate their technical strength in the field of large model training but also provide strong confidence support for the future development of domestic computing power.