Xing4.0-29B-Intent

Headlines · 1

1. 5h
   China Telecom's XingChen lab open-sources Xing4.0-29B-Intent

   On October 19, the XingChen AGI Lab at China Telecom AI open-sourced Xing4.0-29B-Intent, a 29B model that combines intent classification and parameter extraction and returns results in a unified category+args JSON structure. It was released with model weights, a 2.753 million-item SFT corpus and an inference SDK. Zhedongxi reports end-to-end latency under 150 ms and a zero-shot cold-start Micro-F1 of 70.27; it also cites an overall score of 58 on a unified benchmark across 48 public datasets, 9 points above second-placed Gemini-3.6-Flash. In 11 hands-on tests across smart spaces, smart cockpits, enterprise workflows, office collaboration and personal agents, the outlet found solid multi-object parameter binding and multi-task splitting, but flagged an error where "do not open the window" was turned into a 0% opening. The code is open under Apache 2.0 in the TeleChat-USP GitHub repository; the report gives Hugging Face, ModelScope and Modelers links, says datasets and weights are on Hugging Face, and notes an official inference image.

   智东西 · 🔥 9

Experience and discussion from the community

Share my Xing4.0-29B-Intent experienceAsk about Xing4.0-29B-Intent