China Telecom Open-Sources Xing4.0-29B-A4B, Its First Fully Domestic-Stack Lightweight Agent LLM
Summary
China Telecom released and open-sourced its new-generation Xing (Star) model Xing4.0-29B-A4B on September 17, with 29B total parameters, only 4B activated, native 256K context extendable to 512K. It is billed as China's first 10B-plus-parameter model trained entirely on a domestic stack — Ascend training chips, domestic frameworks, in-house architecture — and scored 93.52 on SuperCLUE's agent benchmark, third place and less than one point behind two leading Qwen models. After 4-bit quantization its VRAM footprint drops to 15GB, letting consumer GPUs like the RTX 3090/4090 run long-context tasks locally.
Bubble analysis
This is a model release: it carries no investment figure, valuation or compute purchase, so it offers almost no direct evidence on whether AI is in an investment bubble. Its only indirect relevance is that a fully domestic training stack and local deployment on consumer GPUs suggest model efficiency is improving and the compute bar per model is falling — which over the long run could soften demand for massive capex, but that chain is too long to count as bubble evidence either way.
Related signals
linked by shared tags
- CITIC Securities: Ascend 960 Launch Moved Up, China's Domestic Compute Supply Chain Set to Benefit#ascend
- Huawei's Zhu Zhaosheng: Ascend Has Crossed the Ecosystem Inflection Point and Is Fully Capable of Building a New AI Ecosystem in the Agentic Era#ascend
- A Brief History of Chinese Large Models 03: The Hundred-Model War Begins, Not Every Card Looks the Same#llm
- Alibaba Qwen Open-Sources Qwen-Drive-1.0-4B: First Vision-Language Foundation Model for Autonomous Driving#open-source
- 160,000 Ascend 950DT Chips Deployed: Domestic AI Computing Clusters Enter the 'Gigawatt Era'#ascend