At the prestigious Global 1024 Developers Festival, iFLYTEK made a significant announcement by officially introducing the iFLYTEK Spark X2.6 integrated end-cloud large model. This cutting-edge model represents a groundbreaking achievement, as it is the first to overcome limitations in code and agent capabilities by leveraging domestic computing power. It effectively tackles the challenges of training efficiency for long-duration tasks on locally produced chips, a significant bottleneck in the industry. Moreover, the Spark X2.6 model introduces two pioneering innovations aimed at enhancing inference efficiency for extended tasks within domestic computing clusters.

The Spark X2.6 model is built upon a sophisticated Mixture of Experts (MoE) architecture, boasting an impressive parameter range of 293 billion to 30 billion. It extends the context length to an unprecedented 1 million tokens, setting a new benchmark in the industry for code and agent capabilities. Furthermore, the model supports customizable security policies, providing users with greater control and flexibility. Interested parties can now experience the power of Spark X2.6 firsthand through a trial available on the official iFLYTEK Spark website.

In conjunction with the Spark X2.6 model, iFLYTEK also released the Spark X2.6-4B open-source end-side large model. This model represents a significant leap forward in agent capabilities for end-side large models, delivering substantial improvements across multiple dimensions. Notably, it outperforms ollama in terms of inference efficiency, demonstrating iFLYTEK's commitment to pushing the boundaries of technological innovation. Prior to this release, the Spark X2.5-4B end-side model had already gained international recognition, ranking among the top four globally in terms of downloads following its open-source debut.