iFLYTEK has officially launched the new Spark X2.5 large model. The model core adopts a mixture-of-experts architecture, with a total parameter count of 293 billion, but only activates 30 billion parameters during actual inference, achieving a leap in performance while maintaining efficient operation.

In terms of core capabilities, Spark X2.5 has made significant breakthroughs, particularly enhancing code generation and intelligent agent interaction performance. It also demonstrates stronger overall strength in general areas such as mathematical computation and semantic understanding. In addition, the model fully supports a long context processing capability of 256K and has been deeply optimized for domestic computing power environments, ensuring efficient and stable operation on local computing infrastructure.

In addition to the iteration of the base large model, iFLYTEK has also simultaneously open-sourced two lightweight edge-side models: Spark X2.5-4B and 1.7B. These two edge-side models not only support a context processing capability of up to 1 million tokens, but can also flexibly adapt to various scenarios for local deployment. Currently, the Spark X2.5 series models have been fully launched on the iFLYTEK Open Platform, and related time-limited promotional policies have been introduced, offering developers and enterprise users more cost-effective choices.