iFlytek launches Spark X2.6: a 293B model trained entirely on Chinese-made computing power

#What happened

At its Global 1024 Developer Festival on October 22, 2026, iFlytek launched Spark X2.6, a device-cloud integrated large model. The company says X2.6 was trained entirely on domestic computing power, with a mixture-of-experts architecture of 293 billion total parameters (30 billion activated) and a context window of up to 1 million tokens. Chairman Liu Qingfeng claimed Spark is the only large model trained on domestic chips while others were trained on NVIDIA GPUs. The company also says it broke through the long-context training efficiency bottleneck on domestic chips with dynamic parallel training technology, and reports 4x training and 4.5x inference efficiency gains on Chinese-made hardware. A smaller open-source model, Spark X2.6-4B, was released for on-device use, with improvements across code, agent, long-text, and multilingual benchmarks versus X2.5-4B.

#Why it matters

This is the most concrete demonstration yet that China's frontier models can now train at scale without US chips, a direct answer to export controls. The 293B-A30B MoE size plus a million-token context puts domestic-hardware training in the same league as global frontier models on paper. The domestic-hardware performance claims remain company claims with no independent validation, so treat efficiency numbers as unverified. The aggressive API pricing (0.9 yuan per million input tokens, 0.08 yuan cached, 3.6 yuan output, all limited-time) also signals a price war on Chinese model APIs.

Verified Oct 23 2026