iFlytek Launches Spark X2.6 Model Trained on Domestic Compute
iFlytek released the Spark X2.6 edge-cloud integrated large model on October 22, 2026, marking the first time a domestic LLM has achieved breakthroughs in coding and agent capabilities using entirely domestic computing power. Additionally, the company introduced the Spark X2.6-4B open-source edge model and an edge-cloud collaborative architecture. This launch represents a significant milestone for the domestic AI industry by overcoming hardware and bandwidth bottlenecks without relying on Western GPU hardware like Nvidia chips. It demonstrates that competitive LLM training and efficient edge-cloud integration can be achieved using indigenous hardware ecosystems. The Spark X2.6 model utilizes a 293B-A30B parameter MoE architecture with an expanded 1M context length and competitive pricing for token inference. It also resolves long-range task training and distributed inference bottlenecks specific to domestic heterogeneous computing platforms.
## BACKGROUND
Large language models traditionally rely heavily on high-end foreign graphics processing units and specialized interconnects for training and inference. Achieving breakthroughs on domestic computing power requires overcoming severe memory bandwidth limitations and developing specialized parallel training techniques for alternative hardware.