On September 17, China Telecom proudly announced the launch of its latest innovation, the Xingchen Large Model Xing4.0-29B-A4B. Boasting an impressive total of 29 billion parameters, this model ingeniously activates only 4 billion parameters, ensuring efficiency without compromising performance. It boasts native support for a 256K context window, with the flexibility to extend up to 512K, catering to a wide range of applications.
Notably, the Xing4.0-29B-A4B marks a significant milestone as the first domestic large model with tens of billions of parameters, trained utilizing solely domestic computing power and frameworks. It has undergone profound optimization to excel in complex engineering tasks, achieving full-stack localization from the training chips to the inference deployment stages.
In the rigorous agent capability evaluation conducted by SuperCLUE, the model demonstrated its prowess by securing the third position with a commendable score of 93.52, trailing the top two Qwen models by a mere margin of less than 1 point.
Moreover, following 4-bit quantization, the model's memory consumption is significantly reduced to just 15GB, enabling seamless execution of long-context tasks on 24GB consumer-grade graphics cards, such as the RTX 3090 and RTX 4090, directly on local machines.
Continuing its commitment to open-source principles, the Xing4.0-29B-A4B model ensures full compatibility with mainstream open-source ecosystems and is readily accessible on various open-source communities, fostering collaboration and innovation within the tech community.