At the 2026 iFLYTEK Global 1024 Developer Festival held on October 22, iFLYTEK Chairman Liu Qingfeng publicly discussed the current state of domestic compute and how to break through in the future. He said domestic compute still lags about “nearly two generations” behind NVIDIA, mainly in VRAM and bandwidth. When faced with coding and agent tasks that often require million-token contexts, the industry previously ran into the problem of data that “couldn’t fit and couldn’t be trained.”
To address this challenge, iFLYTEK introduced a distinctive solution with its device-cloud integrated “iFLYTEK Spark X2.6.” The architecture combines a cloud-based large model with a 4B on-device model. Through dynamic parallel training for long-horizon tasks and adaptive compute allocation based on task length, it improves ultra-long context training efficiency by about 4x. On the inference side, thanks to adaptive memory-compute collaboration and the first four-layer KVCache and distributed inference technology based on fully domestic hardware, overall efficiency is improved by 4.5x. Users can decide whether to keep tasks processed locally, use device-cloud collaboration, or send them to the cloud based on task difficulty and data sensitivity.
In building on-device models, iFLYTEK has found an effective path around compute constraints. Although there is an overall compute gap, there is enough compute to train on-device models. Through meticulous refinement, iFLYTEK has raised the completion rate of on-device models on complex agent tasks from the industry’s previous level of just over 12% to 41.9%, and plans to move toward 60% next. On-site demonstrations of scenarios such as business reviews and procurement contract reviews showed that local de-identification and efficient distribution through device-cloud routing not only ensure sensitive data never leaves the device, but also effectively reduce token usage costs. At the same time, Liu Qingfeng also pointed out that in a device-cloud integrated architecture, the cloud’s core role is to provide unified joint training for the device side, assess capabilities, and offer fallback guidance.