On September 7, 2026, iFlytek officially released its Spark X2.5 large model. The model uses a Mixture-of-Experts (MoE) architecture with a parameter size of 293B-A30B, focusing on improvements in code and intelligent agent capabilities while continuing to enhance general capabilities such as mathematics and language understanding.
The model completed the full training and inference pipeline on a fully domestic computing stack, optimizing solutions for ultra-long sequence training and inference challenges on domestic hardware. It has been launched on the iFlytek Open Platform.
iFlytek has also open-sourced two edge-side versions of Spark X2.5, the 4B and 1.7B models, which support up to 1 million tokens of context and can be locally deployed for long-document processing, code assistance, and data analysis. The full Spark X2.5 supports 256K context and covers more than 200 languages.
The release is part of iFlytek's broader push toward full-stack domestic AI capabilities, covering cloud, edge and on-device deployment scenarios.
Source: iFlytek / CNR
