iFlytek Unveils Spark-X2.6: 293B MoE Model Built Entirely on Domestic Chinese Chips
iFlytek open-sources Spark-X2.6, a 293B MoE model with 30B active parameters and 1M context, trained and deployed entirely on domestic 10,000-card chip clusters.
At the 1024 Global Developer Festival in Beijing, iFlytek released and open-sourced Spark-X2.6, a 293-billion-parameter Mixture-of-Experts foundation model trained and deployed entirely on domestic Chinese computing infrastructure. Featuring 30B active parameters per token and a 1-million-token context window, the model represents a breakthrough in long-horizon training on homegrown accelerators.
293-A30B Sparse MoE Architecture with 1 Million Token Context
On October 22-23, 2026, Chinese AI firm iFlytek officially released and open-sourced its flagship general-purpose model, Spark-X2.6, during the iFlytek 1024 Global Developer Festival in Beijing. Spark-X2.6 adopts an end-cloud integrated sparse Mixture-of-Experts (MoE) architecture with 293 billion total parameters and 30 billion active parameters per token. Supporting an ultra-long context window of 1 million tokens, the foundation model delivers advanced agentic planning, mathematical reasoning, and automated code generation for enterprise environments.
Full Independence on Homegrown 10,000-Card Computing Clusters
The landmark technical milestone of Spark-X2.6 is its full development cycle on domestic semiconductor hardware without reliance on restricted Western accelerators. Chairperson Liu Qingfeng emphasized that the model broke through the training efficiency bottlenecks of long-horizon tasks on domestic chips, achieving full algorithmic and architectural independence. Spark-X2.6 is actively deployed across a domestic 10,000-card inference cluster, serving core industrial workloads across healthcare, education, state-owned enterprises, and judicial administrative platforms.
Maturing Ecosystem and the Leap from Inference to Frontier Training
The launch highlights an accelerating alignment between Chinese model developers and domestic chipmakers following tightening export controls. Alongside Spark-X2.6, researchers demonstrated post-training on Huaweis Ascend 910C silicon for the DeepSeek-V4-Pro model, demonstrating that domestic silicon has matured from handling basic inference to executing large-scale, distributed training runs. Industry analysts noted that Chinese AI labs are increasingly prioritizing full-stack domestic hardware integration, data sovereignty, and production deployment resilience over superficial parameter scaling.