Physix Frontier · News Briefing Card (QbitAI · Sep 20, 2026)

China Telecom Open-Sources Xing4.0, Runs on Single RTX 3090

KEY FACTS

China Telecom has open-sourced Xing4.0, a 29B-parameter MoE code-agent large model.
Trained on Huawei Ascend 910C chips, it achieves full-stack domestication from hardware to framework.
After 4-bit quantization, VRAM usage drops to 15GB, enabling local deployment on an RTX 3090.
The model natively supports a 256K context window for long-document processing and complex agent tasks.
It is available on GitHub and Hugging Face, where it reached fourth place in trending repositories.

KEY DATA

29BTotal Parameters

~4BActive Parameters

~15GBPost-Quantization VRAM

~96%Training Throughput Increase

PHYSIX OBSERVATION

China Telecom’s move precisely targets the pain points of data-sensitive enterprises by lowering the barrier to private deployment with consumer-grade GPU requirements. This full-stack domestic adaptation is not just a showcase of technological autonomy; it signals that the Xinchuang market is finally receiving viable lightweight AI office solutions, which could accelerate AI adoption across government and enterprise sectors.

Source: QbitAI report