Acknowledging the lag of domestic computing systems behind leading global counterparts, iFLYTEK presented a strategy during its earnings report on August 21st. Instead of waiting for hardware improvements, the company decided to utilize engineering developments for more efficient use of existing chips.
CEO Liu Qingfeng admitted that the company had been inconspicuous in the recent wave focused on code and agents, partly because the limitations of domestic computing severely restrict the ability to create very long context windows. iFLYTEK notes that its domestic accelerators lag behind Nvidia H200 in training efficiency by about five times, with this gap being most noticeable when training on contexts exceeding 256 thousand tokens.
However, according to Liu Qingfeng, in sectors such as education, healthcare, automotive manufacturing, and state-owned enterprises, 90%–95% of use cases do not require context in millions of tokens. Therefore, the company is optimizing the architecture of models, operators, communications, memory, and frameworks for training. Methods for inference with long sequences have been successfully tested, and a complete domestic chain from chip to model and application has been built.
This approach is reflected in the company's product line: Spark X2 was launched in February, and Spark X2-Flash became the first Mixture-of-Experts model with 30 billion parameters, fully trained and deployed on Huawei Ascend 910B clusters, in April. This was followed by the multimodal Spark X2-VL in June, and a new flagship domestic generative product is expected by the end of August at the 1024 Developer Festival.
The engineering focus extends beyond just models. In July, iFLYTEK received the national first prize for the Pengcheng Cloud Brain project—a large-scale domestic intelligent computing project overseen by the Pengcheng National Laboratory. From a commercial perspective, revenue from LLM API and MaaS grew by approximately 70% year-over-year in the first half. The company employs a strategy of selling results rather than tokens, combining trillion-scale cloud models with on-device models instead of directly competing with the largest cutting-edge model. Liu asserts that this approach helps maintain margin and focus.
Liu believes that iFLYTEK's computational lag behind domestic competitors will equalize within two years—and possibly even turn into an advantage—as chip ecosystems mature.