Xing4.0-29B-A4B: Domestic from Ascend Chips to Frameworks, Consumer-Grade GPUs (www.aibase.com)

China Telecom has announced the Xing4.0-29B-A4B, a next-generation large AI model featuring 29 billion parameters, with only 4 billion activated. Notably, it supports extended contexts of up to 256K tokens, expandable to 512K, and is the first fully domestic large model in China, developed from the ground up with Ascend chips for training and an entirely homegrown software framework. This closed-loop approach eliminates reliance on overseas supply chains, providing a significant boost in technological independence within the AI industry. The significance of this model lies in its high performance combined with accessibility, as evidenced by its impressive SuperCLUE evaluation score of 93.52, ranking it just behind leading models like Qwen. Developers appreciate its efficiency as the model can run with only 15GB of VRAM after 4-bit quantization, making it compatible with consumer-grade GPUs like the RTX 3090 or RTX 4090. This ability allows individuals and small teams to perform complex engineering tasks locally, reducing the need for expensive cloud resources and democratizing access to powerful AI tools.