iFlytek Launches Spark-X2.5 Multilingual MoE Model Trained on Chinese Hardware
iFlytek has officially released the Spark-X2.5 multilingual model, utilizing a 293B-parameter Mixture-of-Experts (MoE) architecture with 30B active parameters and a 256K context window. The model was trained entirely on domestic Chinese computing infrastructure and covers more than 200 languages. This release highlights China's progress in scaling large language models independently on domestic hardware amidst global chip restrictions. Additionally, significant enhancements to its coding, agentic capabilities, and multilingual support make it a competitive alternative for enterprise adoption. The model is deployed on the iFlytek Xingchen Model-as-a-Service (MaaS) platform. Pricing is set at 1.6 RMB per million input tokens (0.24 RMB for cache hits) and 6 RMB per million output tokens.
## BACKGROUND
A Mixture of Experts (MoE) architecture scales model capacity efficiently by routing input tokens only to specific specialized subnetworks (experts), keeping inference compute manageable despite large overall parameter counts. Model-as-a-Service (MaaS) provides developer access to these large models via cloud APIs without requiring localized infrastructure.