At its global 1024 Developer Festival on October 22, iFlytek released Spark X2.6, an end-cloud integrated large model it says is the first trained entirely on domestic compute to break through in coding and agent capabilities. Chairman Liu Qingfeng said Spark is the only major model trained on homegrown chips rather than Nvidia GPUs, and at the smallest size among its peers.
X2.6 uses a 293B-A30B mixture-of-experts architecture, extends context length to 1 million tokens, reaches first-class coding and agent capability, and offers task-customizable security policies. Inference pricing is set at 0.9 yuan per million input tokens (0.08 cached) and 3.6 yuan per million output tokens. iFlytek said it solved the long-task training-efficiency bottleneck on domestic chips with dynamic parallel training and distributed inference tailored to local heterogeneous hardware.
The company also released the open-source X2.6-4B edge model, the first on-device model with a leap in agent ability; since the X2.5-4B edge model opened in September its downloads have entered the global top four for edge models. On stage iFlytek also showed a domestic voice base model, Spark-Audio, the WinWork digital-employee platform, AI glasses, an AI recording ring and a robot super-brain platform.
Source: iFlytek 1024 Developer Festival via IT Home / Tencent News, October 22, 2026
