iFLYTEK unveils Spark X2.5 with moe architecture and enhanced code capabilities

iFLYTEK has officially launched the Spark X2.5 large model, which employs a MoE architecture with a parameter scale of 293B-A30B. The model focuses on improving code and agent capabilities, as well as general skills like mathematics and comprehension. The entire training and inference process utilized fully domestic computing resources, addressing challenges in ultra-long sequence training and inference within the country. The model is now accessible on the iFLYTEK Open Platform. Previously, edge-side models of Spark X2.5, including versions with 4B and 1.7B parameters, were released as open source, supporting up to 1 million Tokens. These models can be deployed locally for tasks such as long document processing, code assistance, and data analysis. With the release of the X2.5 flagship model, iFLYTEK has completed its new version layout across both cloud and edge product lines.