China's iFlytek Launches Spark X2.6 Trained Entirely on Domestic Chips
Chinese artificial intelligence company iFlytek used its annual 1024 Developer Festival on Thursday to launch Spark X2.6, an end-cloud large model that it says is the first to achieve breakthroughs in both coding and agent capabilities while trained entirely on domestic computing power. Chairman Liu Qingfeng told the audience that among companies building large models, iFlytek is the only one training purely on Chinese-made chips, while almost all others run on Nvidia hardware. Its models, he added, are also the smallest in size among the major players.
The end-cloud label describes an architecture in which heavy computation happens in data centers while lighter tasks run on devices, a design meant to keep services responsive and reduce the cost of serving millions of users. For iFlytek, the pitch is less about raw scale than about squeezing useful performance out of constrained hardware.
The constraint has a name: memory bandwidth. Domestic AI chips generally offer less of it than the top Nvidia accelerators, which slows the long chains of reasoning that modern agents depend on. iFlytek said Spark X2.6 was engineered around this limitation, using long-horizon task techniques to compensate for the narrower data pipe. The company did not disclose benchmark scores or parameter counts in its presentation.
The stakes go beyond one product. China’s AI industry has spent years building alternatives to Nvidia’s CUDA software ecosystem and its chips, and a model that performs well on domestic silicon offers a proof point for that effort. If coding assistants and agents can run reliably on Chinese hardware, the argument that the country’s AI progress is capped by access to foreign processors weakens considerably.
The 1024 Developer Festival, named for the 1,024 bytes in a kilobyte, is iFlytek’s annual showcase and has become a venue for the Hefei-based company to position itself as a full-stack player, from chips and models to consumer devices such as translators and voice recorders. This year’s event put domestic computing front and center.
Competition is intense. DeepSeek, Alibaba and ByteDance all field capable models, and several have emphasized efficiency and low inference costs over sheer size. iFlytek’s differentiator is the claim of a fully domestic stack, which carries weight with enterprise buyers in sectors that prefer homegrown technology.
What remains unproven is performance. Announcing capabilities in coding and agents is not the same as matching the best models on independent tests, and iFlytek has not published comparisons. Developers will be the judge: the model’s reception on coding tasks and agent workflows in the coming weeks will show whether a domestic-chip model can compete on merit rather than provenance.