- China Telecom's XingChen lab open-sourced Xing4.0-29B-Intent, which returns a request's intent and its exact parameters together in one JSON output.
- The team reports strong benchmark scores, but these are its own results. One hands-on test published by Zhidx showed a car-window error.
- Before any agent acts on such output, teams should check the parameters, not just the intent, against their own data.
The risk moves from what the AI says to what it fills in
An AI assistant that gives a wrong answer is an annoyance. An assistant that writes a wrong value into a command is a different problem, because a system then acts on it. This is the shift behind a new open release from China Telecom.
This shows how the risk changes as agents move from talking to acting. The number in the field matters more than the sentence around it. Accuracy on settings becomes the measure that counts.
What China Telecom released
China Telecom's XingChen AGI lab open-sourced Xing4.0-29B-Intent, according to Pandaily. The model has one narrow job. Given a user's request, it identifies the action wanted and also supplies the specific settings the receiving system needs.
Pandaily notes that many intent models stop at classification. They can tell a user wants to adjust the lights. They cannot say which lights, to what level, or which to leave alone.
The lab started from its XingChen 29B base model and tuned it for reading meaning and producing structured results. The release also includes a training dataset and an inference SDK.
How it works
The model is given a schema. A schema is a template that lists the actions a system supports and the fields each action needs. The model returns one JSON output, a structured text format that software can read directly.
That output has two parts. A category field holds the detected intent. An args field holds the extracted entities, values and attributes. If one sentence contains several commands, the model returns several intents, each with its own arguments.
The team says the output is constrained by the schema, which keeps formatting errors low. It reports end-to-end latency under 150 milliseconds. Doing both steps in one pass means no hand-off between a classifier and a separate extractor.
What the team reports, and what it does not prove
In a zero-shot cold-start test, the team scored 70.27 on Micro-F1 and says that beats GPT-4.1. Its exact-match result was 61. The source does not say what that exact-match figure measures. Zero-shot means the model had no examples of the task beforehand.
A second, broader test mixes four language tasks: finding named things in text, linking the relations between them, picking out events, and understanding spoken requests. The team puts its overall result at 58. Gemini-3.6-Flash came second, nine points behind. Pandaily states these are the team's own results.
Benchmarks do not use your schema, your devices or your customers' phrasing. They show a model can do the task. They do not show it will do it for you.
Hands-on tests show both strengths and a slip
Zhidx published hands-on tests, as Pandaily reports. From a single command, the model produced separate settings for different lights. It left alone devices the user had excluded. It held a meeting at its original 30 minutes across several rounds of edits.
It also did not treat instructions written inside a memo as new tasks. That matters because agents read text they did not write, such as emails and documents. The tests do not say how the model handles deliberate attacks.
It stumbled once. In a car-window test, the model returned a 0% setting for rear windows that could change from where they already were.
This is our interpretation. A value like this would still be valid JSON, which is why it could slip past a format check.
The release goes beyond the weights
The training set holds 2.753 million supervised fine-tuning samples, reworked from 61 public datasets. They span 19 domains and 52 languages. EMNLP 2026 has also accepted a related project focused on intent data for phone interactions, the team says.
The source code lives in the TeleChat-USP project on GitHub and carries an Apache 2.0 license. The model and dataset can be downloaded from Hugging Face and ModelScope. An inference SDK and an official image ship alongside them.
The team aims the model at connected homes and buildings, in-car assistants, office workflows and personal AI agents.
What leaders should ask their teams
If your products or workflows plan to let an AI trigger actions, put these questions to the team that owns them.
First, do we measure parameter accuracy, or only whether the intent was right? As general advice, ask for exact match on arguments. A wrong value there sits closest to real harm.
Second, what checks sit between the model's output and the action? Range limits and allowed-value lists can catch values that are well formed but wrong.
Third, what happens when the user did not specify a value? A model that fills a gap with a default has made a decision for the user.
Fourth, is text from documents and messages kept apart from user commands? In one hands-on test the model did not treat memo text as a command. Check your own pipeline.
Finally, test any open model on your own schema before you trust a published score.
The lesson of this release is that fluent output is no longer the hard part. The hard part is a correct value in the right field, every time it matters.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Pandaily.
