China Telecom's Xingchen lab open-sources Xing4.0-29B-Intent for intent plus slot extraction

China Telecom AI's Xingchen general AI lab released and open-sourced Xing4.0-29B-Intent, a model built specifically for intent understanding, on October 22, publishing 2.75 million training examples, the model weights and an inference image alongside it. According to QbitAI, the goal is to turn a single spoken sentence into structured information a business system can act on.

Unlike a general chat model, it does two jobs at once. One is multiple choice: deciding what the user wants from a fixed set of candidates, such as controlling a light. The other is fill in the blank: extracting the details needed to finish the task, such as the room being the bedroom, the device being the light, and the target state being warm yellow. The company says it handles sentences carrying multiple intents, parameters and entities.

The implementation uses what the team calls a unified schema-driven semantic parsing framework, organizing output as a hierarchy of categories and parameters and returning JSON under predefined field rules. Intent classification, entity recognition, relation extraction and event extraction are forced into one input-output format so downstream systems can integrate more predictably and migrate with less work.

Why not just use a general model? The post gives three reasons: latency, cost and reliability. General models tend to reason at length, which a ticket-routing query cannot wait for; at millions of calls a day, using a large model for classification and extraction multiplies cost; and free-text answers are hard to parse stably, while the specialized model's fields and format can be validated.

The team reports two internal evaluations. The first sampled 6,000 questions from 48 public Chinese and English datasets across four extraction tasks and scored both category and content correctness, where Xing4.0-29B-Intent totaled 58 against a best of 49 among the 18 general models tested. The second tested category decisions only, including single-answer, multi-answer and deliberately empty negative cases, and the model scored 70 overall, which the company says beats the comparison models on all three sub-scores.

Those evaluations were run by the company's own team, and the post notes that results reflect only this data and setup rather than all tasks or real deployments. That caveat matters when judging how good the model actually is.

Deployment is already under way: the model handles intent recognition and intelligent ticket routing in China Telecom's customer service system, and converts spoken commands into structured device instructions in smart speakers. The open-source package includes 2.75 million supervised fine-tuning examples rebuilt from 61 public datasets, weights, an inference image and an open schema.

The value of a specialized small model is not that it is smarter, but that it pins down one otherwise unstable link in the chain. For teams building device interaction, enterprise service bots or agents that need function-call arguments, a constrained semantic entry layer answering within 200 milliseconds is often more useful than a stronger model whose output is free-form. What to watch next is the quality of the released corpus, long-tail intent behavior after community fine-tuning, and stability on compound commands.