Xing4.0-29B-A4B: China Telecom's new 31B-parameter MoE model

On September 16, the model XingChen-AGI/Xing4.0-29B-A4B was published on HuggingFace — a text-generation model from the Xing series (formerly TeleChat), developed by China Telecom Artificial Intelligence Technology. With 29 billion total parameters, only 4 billion are activated per token, and the context is 256K with extension up to 512K. The model was trained entirely on the Ascend NPU platform with the MindSpore framework.

What it means

The full weights in safetensors take up 58.1 GB, with the largest file at 4.1 GB. The license is apache-2.0. It runs via transformers, and compatibility with vLLM, SGLang and KTransformers is also stated. There are three quantized GGUF builds on the site, the most compact working one being Q8_0 at 12 GB from jmarceno. Running the full model locally will require a substantial amount of VRAM, but the Q8_0 build at 12 GB looks realistic for consumer GPUs. The code has been released as open source, so a build for your own hardware can be reproduced if desired. For more news on local AI tools, see the CPU3D news feed.