China Telecom open-sources AI model that turns commands into settings

A single spoken sentence can hide several orders: change the living-room light, leave the bedroom alone, and move a meeting. Until now, most systems that understood such sentences had to be stitched together from separate tools. The XingChen AGI lab at China Telecom Artificial Intelligence Technology has open-sourced Xing4.0-29B-Intent, a model that tries to do the whole job in one step.

The usual approach is to classify first and extract later. Many intent models stop at classification: they know the user wants to adjust the lights, but not which lights or to what level. Xing4.0-29B-Intent combines both steps. Given a schema, it returns a JSON structure with a category field for each detected intent and an args field for the entities, values and attributes it extracted. One sentence with several commands produces several intents, each carrying its own arguments.

The team's own benchmarks are the strongest claims. It 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 covering 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 numbers have not yet been reproduced by outside teams. The hands-on tests published by Zhidx show the model working well on most cases but failing on at least one: a car-window command that returned a 0% setting for rear windows that could change their existing position.

The release goes beyond the weights. Training used 2.753 million 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 phone-interaction intent data was accepted to EMNLP 2026. The code is on GitHub under the TeleChat-USP project with an Apache 2.0 license, and the model and dataset are on Hugging Face and ModelScope, with an inference SDK and an official image.

For developers building agents, the gain is practical: a model that returns structured commands in one pass, with reported latency under 150 milliseconds, can be downloaded and tested today. The target uses named by the team are smart spaces, in-car assistants, office workflows and personal agents. The sources do not give a commercial timeline or any deployment figures, so what changes for end users depends on how many products pick it up and how it performs outside the team's own tests.