The competition for computing power in the field of artificial intelligence is moving toward a more practical "localized" deep area. At the recent Yangtze River Delta Robot and Automation Exhibition and the Wuxi Embodied Intelligence Robot Industry Chain Partner Conference, iFLYTEK officially launched the new Starfire multimodal large model X2-VL. The release of this model is not only a technical iteration but has also attracted widespread attention in the industry due to its status as "the only one trained on fully localized computing power."

According to iFLYTEK, the Starfire X2-VL uses a specifically designed MoE (Mixture of Experts) architecture to improve the model's processing efficiency in complex tasks. More importantly, the model was trained entirely on the "Taihu Star Leap" computing platform located in Wuxi. To ensure the continuity of technology implementation and subsequent in-depth operations, iFLYTEK has established a dedicated subsidiary in the Wuxi High-tech Zone, using this model as a core driver for local industrial upgrading.

In terms of actual performance, Starfire X2-VL has delivered a competitive result. For mixed text and image questions covering multiple subjects at the high school level, its accuracy rate has approached 95%. In a more rigorous third-party test, this large model directly challenged the 2026 National Mathematics I Paper of the Gaokao. Under the evaluation of an expert group composed of two national-level math teachers, it scored 148 points, demonstrating strong logical reasoning and graphic analysis capabilities.

As AI application scenarios shift from "exploration" to "productivity," how to achieve a leap in model capabilities while ensuring the autonomy and controllability of computing power has become the focus of competition among large model manufacturers. The successful practice of Starfire X2-VL on the domestic computing infrastructure undoubtedly provides an extremely valuable reference model for the industry, indicating that domestic AI models are gradually forging an independent and accelerated path in complex application implementations.