(This article is authored by TechPulse and published with authorization from Titanium Media.)
“Domestic computing power still lags behind NVIDIA by nearly two generations today, though we’re catching up rapidly,” said Liu Qingfeng, Chairman of iFLYTEK, at the 2026 iFLYTEK Global 1024 Developer Festival on October 22. “However, domestic computing power is improving, and I believe it will close the gap within the next two years.”
At this conference, iFLYTEK unveiled the Xunfei Spark X2.6, an integrated edge-cloud solution featuring a cloud-based large model paired with a 4B on-device model. Whether a task runs locally, leverages edge–cloud collaboration, or is offloaded to the cloud is determined by the task’s complexity and data sensitivity.
iFLYTEK Spark X2.6 employs fully domestically produced computing power for training; however, domestic computing capabilities still lag behind NVIDIA’s by roughly “two generations,” according to Liu Qingfeng. The gap is primarily in GPU memory and bandwidth. In this wave, code and agent-based tasks often require contexts on the order of one million tokens, a situation he summarizes as “unable to fit, unable to train.”
iFLYTEK’s solution is dynamic parallel training for long-range tasks, which adaptively allocates computing resources based on task length, boosting ultra-long-context training efficiency by approximately fourfold. On the inference side, leveraging adaptive storage–compute coordination, combined with the first domestically produced four-layer KVCache and distributed inference technologies, performance is improved by a factor of 4.5.
Edge models, first find a way to break through.
“What truly bypasses current computational‑power constraints are edge‑side models,” said Liu Qingfeng. “Although our computing power still lags behind that of the United States, we have sufficient resources to train edge‑side models.” Once the models are scaled down, the computational‑power bottleneck is largely mitigated, enabling us to compete on equal footing with overseas counterparts at this level.
How much work the edge side can handle depends on the completion rate of complex intelligent-agent tasks.
According to reports, previously, edge‑side models in the industry achieved only about 12.3% accuracy on this type of agent‑based task, whereas iFLYTEK’s 4B model has boosted that figure to 41.9%, with a next‑step target of 60%. Smartphones, PCs, AI glasses, intelligent cockpits, and robots all fall under the edge‑side category, and routing underpins this division of labor; iFLYTEK reports an edge‑cloud routing accuracy of 96.1%.
Two scenarios were demonstrated on-site. One was a business review for the “CEO of a robotics startup,” where the model locally correlated weekly reports, purchase notifications, and project emails to identify recurring follow-ups on the same customer, with all data remaining within the local environment. The other scenario involved reviewing procurement contracts: sensitive information was first anonymized locally and sent only after user confirmation, while the original local files were never uploaded.
This approach entails a clear trade-off: leveraging edge-cloud integration to address the challenges of deploying large models in a way that is both secure and cost-effective in terms of token usage. Liu Qingfeng also emphasized, “Since domestic computing power is still limited, our strategy is to combine cloud and edge; we won’t build overly large models for the cloud either.”
Meanwhile, in this edge-cloud integrated architecture, the cloud’s role is more than just a fallback.
Liu Qingfeng stated that if you want to train on-device models, you can’t do it effectively without the capabilities of cloud-based models; whether the on-device tasks meet the required standards must be evaluated using large cloud models. Moreover, the boundaries of on-device capabilities need to be calibrated by the cloud. This time, iFLYTEK is implementing joint, unified training across the cloud and edge.
In his view, vendors that focus solely on edge‑side models lack the fundamental capabilities of cloud‑based large models; they cannot effectively guide the edge and do not know which tasks should remain on the edge and which must be offloaded to the cloud, making it highly unlikely that they will have a competitive advantage.
Calling on more vendors to train large models using domestically produced computing power.
In addition to the end-cloud integrated iFLYTEK Spark X2.6, iFLYTEK also unveiled the SparkAudio 1.0 large-scale speech foundation model, the Spark Embodied Multimodal Interaction technology and related products, as well as the iFLYTEK Xingfan WinWork digital employee platform and the iFLYTEK AIPC product series. Furthermore, it released new technical solutions for developers in fields such as smart hardware and enterprise applications.
Liu Qingfeng stated, “We don’t want iFLYTEK to be the only one conducting training on domestically produced computing power; by speaking out loudly, we hope more players will join in.”
“Right now, only Huawei and we are the main players,” he explained. “With NVIDIA, algorithmic innovations are open-source, so everyone can use them directly—just grab the operators and code and run it. But on domestic computing platforms, the architectures are different, and many pitfalls remain unexplored. When it comes to pushing models to their limits on these domestic systems, tackling the problem collectively as ten companies versus doing it alone results in a cost difference of an order of magnitude.”
"If everyone uses domestically produced computing power, what is there to fear?" Liu Qingfeng said. "We’re armed with a rifle and a few sticks of dynamite, while they already have planes and artillery."
Regarding the pace at which the gap is narrowing, he predicts that it will level out over the next two years—domestic computing power is improving, while the U.S. is imposing restrictions. He also touched on a more distant horizon: iFLYTEK is already laying the groundwork for quantum computing. If, within three to five years, quantum computing can be integrated with today’s classical AI pre-training methods, the fundamental issue of computational scale could be completely redefined.
Under current computing power constraints, he outlined three strategic positioning areas for iFLYTEK: benchmarking against the world’s top models on a fully stack‑independent and controllable platform, with the goal of keeping pace; achieving global leadership in sectors such as education, healthcare, energy, and security; and offering a multilingual large model that positions China as a second choice for the global market. (This article was first published on the TMTPost app; author | TechPulse.)
For more exciting content, follow TMTPost on WeChat (ID: taimeiti), or download the TMTPost App.