At the 2026 iFLYTEK Global 1024 Developer Festival conference held on October 22, iFLYTEK Chairman Liu Qingfeng publicly discussed the current state of domestic computing power and the future path to breaking through. He said that domestic computing power currently lags behind NVIDIA by roughly 'nearly two generations,' mainly in memory and bandwidth. When facing code and agent tasks that often require a million-token context, they previously faced the predicament of 'can't fit it, can't train it.'
To meet this challenge, the device-cloud integrated 'iFlytek Spark X2.6' released by iFLYTEK this time offers a unique solution. The architecture uses a combination of one cloud large model paired with one 4B on-device model, through long-horizon task dynamic parallel training, adaptively allocating computing power according to task length, boosting ultra-long-context training efficiency by about 4 times. On the inference side, relying on adaptive memory-compute collaboration and the first four-layer KVCache and distributed inference technology based entirely on domestic equipment, overall efficiency was improved by 4.5 times. Users can decide autonomously, based on task difficulty and data sensitivity, whether to keep a task local, coordinate between both ends, or hand it to the cloud.
In building on-device models, iFLYTEK has found an effective path around the computing power constraint. Although overall computing power has a gap, the compute for training on-device models is sufficient. Through fine polishing, iFLYTEK raised the on-device model's completion rate on complex agent tasks from the industry's previous twelve point something percent sharply up to 41.9%, and plans to move toward 60% next. Live demonstrations of scenarios such as business review retrospectives and procurement contract audits showed that through local desensitization and efficient device-cloud routing distribution, sensitive data can be kept on-device while also effectively reducing Token usage costs. At the same time, Liu Qingfeng also noted that in the device-cloud integrated architecture, the core role of the cloud lies in joint unified training of the device side, capability assessment, and fallback guidance.
When discussing the domestic computing power ecosystem, Liu Qingfeng called on more vendors to join the ranks of training large models on domestic computing power. He admitted that currently only a few companies such as Huawei are pushing it forward as a main effort; because the domestic computing architecture differs from NVIDIA's and 'no one has stepped in the pits before,' the cost of going it alone is extremely high, but if everyone works together, the shared costs would fall by orders of magnitude. On the speed of narrowing the gap, he judged that computing power will gradually level up over the next two years, and iFLYTEK is also actively laying out quantum computing, aiming to fundamentally rewrite the magnitude problem of computing power within three to five years.
Under current computing power constraints, iFLYTEK has clarified three strategic positions: benchmark against the world's top models on a fully self-controlled, autonomous platform and follow closely; achieve global leadership in vertical industries such as education, healthcare, energy, and security; and, with multilingual large models, represent China in offering the world a second choice.