iFLYTEK has proudly announced the official release of its Spark X2.5 large-scale model. This cutting-edge model leverages a Mixture of Experts (MoE) architecture, boasting a parameter range that spans from an impressive 293 billion (293B) to a still-substantial 30 billion (30B). A key focus of the Spark X2.5 model lies in significantly enhancing its code and agent capabilities. Simultaneously, it is committed to the ongoing refinement of general-purpose abilities, including mathematical prowess and comprehension skills.

Notably, Spark X2.5 has accomplished its entire training and inference journey utilizing solely domestic computing resources. This achievement effectively overcomes the hurdles associated with training and inference for ultra-long sequences within the Chinese context. The model is now readily accessible on the iFLYTEK Open Platform, opening doors to a wide range of applications.

Prior to this, iFLYTEK had already open-sourced two edge-side models, namely Spark X2.5-4B and 1.7B. These models are designed to support contexts of up to 1 million tokens, facilitating local deployment in various scenarios. These include long document processing, code assistance, and data analysis, thereby catering to the diverse needs of users in these domains.