AI BriefingXing4.0-29B-A4B: 29B MoE Model Trained on Ascend NPU, with Only 4B Parameters Active
XingChen-AGI/Xing4.0-29B-A4B
About the project
The model features an MoE architecture that activates only 4B parameters per token out of a total of 29B, with a base context of 256K that can be extended to 512K. It is the latest model in the Xing series developed by China Telecom AI, and is notably trained entirely on the Ascend NPU platform and the MindSpore framework.
Based on mHC, MLA, and MTP architectures, it is optimized for agent tasks. It reliably performs multi-step planning, tool calling, and complex reasoning chains even with long contexts, and integrates smoothly with agent frameworks such as OpenCode and Claude Code.
It supports major inference frameworks including vLLM, SGLang, and KTransformers, and allows fine-tuning via LLaMA-Factory. It is designed to facilitate lightweight customization for specific domain tasks such as contract auditing and table understanding.
XingChen-AGI/Xing4.0-29B-A4B
The original page has no description.
text-generation
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.