The XingChen AGI lab at China Telecom Artificial Intelligence Technology has open-sourced Xing4.0-29B-Intent, a model built for one narrow but critical job in AI agents: working out what a user wants and then filling in the exact parameters a system needs to act on it.
Many intent models stop at classification. They can tell that a user wants to adjust the lights, but not which lights, to what level, or which ones to leave alone. Xing4.0-29B-Intent combines intent classification and parameter extraction in a single pass. Given a schema, it returns a unified JSON structure with a category field for the detected intent and an args field holding the extracted entities, values and attributes. One sentence that contains several commands produces several intents, each with its own arguments.
The model is fine-tuned from the XingChen 29B base model and optimized for semantic parsing and structured output. The team reports end-to-end inference latency under 150 milliseconds, with schema-constrained output that keeps formatting errors low.
On benchmarks, the team reports a zero-shot cold-start Micro-F1 of 70.27, which it says beats GPT-4.1, and an exact-match score of 61. On a broader evaluation spanning named entity recognition, relation extraction, event extraction and spoken language understanding, it reports an overall score of 58, nine points above Gemini-3.6-Flash in second place. These are the team's own results.
In hands-on tests published by Zhidx, the model separated settings for different lights in one command, skipped devices the user said not to touch, kept a meeting's original 30-minute duration through several rounds of changes, and did not treat instructions written inside a memo as new tasks. It also stumbled in one car-window test, returning a 0% setting for rear windows that could change their existing position.
The release goes beyond weights. Training used 2.753 million supervised fine-tuning samples rebuilt from 61 public datasets across 19 domains and 52 languages, which the team calls the largest open intent-understanding corpus. Related work on intent data for phone interactions has been accepted to EMNLP 2026. Code is on GitHub under the TeleChat-USP project with an Apache 2.0 license, while the model and dataset are on Hugging Face and ModelScope with an inference SDK and official image.
Target uses include smart spaces, in-car assistants, office workflows and personal agents.