Buyers who want AI built without Nvidia hardware now have a low-priced flagship to test, though iFlytek concedes Chinese chips trail global ones by about two generations.
iFlytek has launched Spark X2.6, a large language model it says was trained entirely on Chinese-made computing hardware, at its annual 1024 Developer Festival on October 22. The model has stronger coding and AI-agent abilities than its predecessor, according to IT Home, which reported a mixture-of-experts design with 293 billion parameters, about 30 billion of them active for each task, and a context window extended to 1 million tokens.
Chairman Liu Qingfeng told the audience that among the mainstream models iFlytek compared itself with, Spark was the only one trained on domestic compute, while the others used Nvidia GPUs. He also said Spark had the smallest parameter count of the group, IT Home reported. iFlytek did not name the comparison models in text, and its benchmark charts were shown only as images.
The launch matters less for its scores than for its supply chain. With X2.6, iFlytek is pitching its reliance on Chinese chips to customers as a feature: control over the full stack, from chips to deployment, for government, education and enterprise buyers worried about supply security.
iFlytek claims efficiency gains on domestic clusters
According to China Fund News and Jiemian, iFlytek said dynamic parallel training, which allocates compute according to task length, made very long-context training about four times as efficient on domestic clusters. It said a combination of adaptive inference, a four-layer cache and distributed inference raised inference efficiency 4.5 times. None of the outlets named the chips or the size of the training cluster, and the claims have not been independently tested.
Liu framed those gains as workarounds for weaker hardware. iFlytek’s chairman Liu Qingfeng said domestic computing power still trails the global level by about two generations. He added that he expects the gap to narrow, China Fund News reported. He said security and cost-effectiveness for large models have to be solved by combining devices and the cloud.
That device-cloud design runs through the launch. iFlytek said X2.6 routes tasks between on-device and cloud models with 96.1% accuracy, Beijing Business Today reported. Liu forecast that about 60% of application scenarios will run on devices, 30% through device-cloud collaboration and 10% on cloud models alone.
Low prices and an open edge model aim at developers
IT Home reported limited-time API prices of RMB0.9 per million input tokens, RMB0.08 per million cached input tokens and RMB3.6 per million output tokens. iFlytek also open-sourced an edge version, Spark X2.6-4B. A model card on Hugging Face lists the 4-billion-parameter model under an Apache 2.0 license, with three reasoning levels and support for up to 1,048,576 tokens of context. The card credits the "SparkLLM Team" rather than naming iFlytek; the same account hosts the X2.5 models iFlytek open-sourced in September.
Some of the headline specifications are not new. IT Home reported the same 293B-A30B architecture for Spark X2.5 in September, which suggests X2.6 refines that base model rather than replacing it. The parameter count, context length and prices all come from IT Home; the other outlets at the event did not report them.
Alongside the model, iFlytek launched a speech foundation model, SparkAudio 1.0, and WinWork, a platform for building AI "digital employees" for business tasks, China Fund News reported.
Results show slow growth while AI spending stays heavy
The launch follows iFlytek’s third-quarter results, released on October 19. Revenue for the first nine months rose 7.05% to RMB18.187 billion and net profit attributable to shareholders was RMB21.89 million, up 132.83%, Yicai reported. In the third quarter alone, net profit was RMB226 million, but excluding one-off items the company lost RMB126 million. Third-quarter R&D spending exceeded RMB1.7 billion, more than a quarter of revenue.
Those numbers make pricing the key question for X2.6. iFlytek is spending heavily to stay on a domestic-only path, and its profit depends on whether low API prices and on-device deployment bring paying customers. The next signals are whether rival developers publish their own results on domestic-only training, and whether X2.6 shows up in iFlytek’s fourth-quarter enterprise and government revenue.