iFlytek releases the device-cloud integrated iFlytek Spark X2.6, trained on fully domestic computing power and paired with a 4B on-device model

"Domestic computing power still lags nearly two generations behind Nvidia today, though we are catching up fast." Liu Qingfeng, chairman of iFlytek, said this on October 22 at the 2026 iFlytek Global 1024 Developer Festival. "But domestic computing power is improving, and I think it will catch up within the next two years."

At this conference, iFlytek released the device-cloud integrated iFlytek Spark X2.6, a cloud large model paired with a 4B on-device model, which decides based on task difficulty and data sensitivity whether a task stays local, is handled jointly by both ends, or is sent to the cloud.

iFlytek Spark X2.6 was trained on fully domestic computing power. Domestic computing power still has a gap with Nvidia, and Liu Qingfeng's assessment was "nearly two generations." The gap is mainly in memory and bandwidth. This wave of code and agent tasks often requires a one-million-token context, and he summed up the previous state as "can't fit it, can't train it."

iFlytek's solution is dynamic parallel training for long-horizon tasks, adaptively allocating computing power according to task length, improving training efficiency for ultra-long contexts by about 4 times; on the inference side, it relies on adaptive storage-compute collaboration, plus the first four-layer KVCache based on fully domestic equipment and distributed inference technology, improving efficiency by 4.5 times.

On-device models: finding a way to break through first

What truly bypasses current computing power constraints is the on-device model. Liu Qingfeng said, "Although our computing power still lags behind the United States, we have enough computing power to train on-device models." Once models get smaller, computing power limits are largely offset, allowing competition with overseas players at this level.

How much work the device side can take on depends on the completion rate of complex agent tasks.

According to the introduction, previously the industry's on-device models had a completion rate of only just over 12 percent on such agent tasks, and iFlytek's 4B model pushed it to 41.9%, with the next goal being 60%. Phones, PCs, AI glasses, smart cockpits, and robots all count as the device side. Supporting this division of labor is routing, and iFlytek's device-cloud routing accuracy figure is 96.1%.

Two scenarios were demonstrated on site. One was a "robot startup CEO" business review, where the model locally linked weekly reports, procurement notices, and project emails, identifying the problem of repeatedly following up with the same customer, with the materials never leaving the device. The other was procurement contract review, where sensitive information is first de-identified locally, sent only after user confirmation, and the original local files are not uploaded.

This approach corresponds to a clear trade-off: using device-cloud integration to solve the deployment problem of large models being both secure and token-cost-saving. Liu Qingfeng also emphasized, "Because domestic computing power itself is currently limited, our positioning is to combine cloud and device, and we won't make the cloud model too large either."

At the same time, in this device-cloud integrated architecture, the cloud's role is not just a fallback.

Liu Qingfeng said that if you want to train on-device models, you can't do it well without cloud model capabilities; whether what the device side does meets the standard needs to be judged by the cloud large model; and where the device's capability boundary lies needs to be calibrated by the cloud. What iFlytek did this time is joint unified training of cloud and device.

In his view, vendors that only make on-device models, lacking the basic capabilities of cloud large models, cannot guide the device side and do not know which tasks should stay on the device and which must go to the cloud, and most likely do not have a competitive advantage.

Calling on more vendors to train large models with domestic computing power

In addition to the device-cloud integrated iFlytek Spark X2.6, iFlytek also released the Spark speech foundation large model SparkAudio 1.0, Spark embodied multimodal interaction technology and related products, as well as the iFlytek Xingfan WinWork digital employee platform and the iFlytek AIPC series products, and released new technical solutions for developers in fields such as smart hardware and enterprise applications.

Liu Qingfeng said, "We don't want iFlytek to be the only one training on domestic computing power; we are loudly calling for more people to join."

"Right now only Huawei and us are doing it as major players." He explained that algorithmic innovations on Nvidia, once open-sourced, can be used directly by everyone; operators and code can run as soon as they are taken; domestic computing power has a different architecture, many pitfalls have not been stepped on by anyone, and for the problem of pushing models to their limits on domestic computing power, ten companies working together versus one company working alone differs in shared cost by an order of magnitude.

"If everyone uses domestic computing power, what do I have to fear?" Liu Qingfeng said. "We are like millet plus rifles, while others already have planes and cannons."

As for the speed at which the gap will narrow, his judgment is that it will close within the next two years, as domestic computing power itself is improving while the United States is restricting. He also mentioned a further layer: iFlytek is already laying out quantum computing, and if within three to five years quantum computing can be combined with current classical AI pre-training methods, the magnitude problem of computing power will be fundamentally rewritten.

Under current computing power constraints, he set three positions for iFlytek: benchmark against the world's top models on a fully self-controllable stack, with the goal of closely following them; achieve global leadership in industries such as education, healthcare, energy, and security; and provide the world with a second option through multilingual large models on behalf of China. (This article was first published on TMTPost APP, by TechPulse)