China Telecom Xingchen Releases and Open-Sources Intent Understanding Model Xing4.0-29B-Intent

China Telecom Artificial Intelligence Technology Co., Ltd.'s Xingchen General Artificial Intelligence Laboratory has officially released and open-sourced Xing4.0-29B-Intent, a dedicated intent understanding model, and simultaneously opened 2.75 million training corpus entries, model weights, and inference images. The model combines intent recognition with parameter extraction, converting a user's single sentence into structured information that business systems can continue to process.

At present, the capabilities of the Xing4.0-29B-Intent model have begun to be applied in terminal products such as China Telecom's 10,000-line customer service and smart home speakers, serving dialogue intent recognition, intelligent ticket routing, and spoken command parsing. This open-source release will provide more developers with the model and supporting resources, supporting integration into their respective business scenarios.

"Help me set the bedroom light to warm yellow." A daily device operation command contains both what the user wants to do and the specific information needed to complete the operation. The system needs to extract this information from natural expressions before it can hand it over to downstream applications for processing.

Faced with this sentence, merely determining "control lighting" is not enough. The system also needs to know: which room? Bedroom; what device? Light; set to what state? Warm yellow.

Jev-type decision models mainly "do multiple-choice questions," returning typed decisions and selecting answers from given candidates; parameter extraction is equivalent to "doing fill-in-the-blank questions," extracting the specific information needed to complete a task. Xing4.0-29B-Intent supports both types of tasks simultaneously, and also supports situations where a single sentence contains multiple intents, multiple parameters, and multiple entities.

The model adopts the Unified Schema-driven Semantic Parsing framework (USP), organizing results in a hierarchical structure of "category + parameters" and outputting JSON-format information according to predefined field rules. Heterogeneous task objectives such as intent understanding, entity recognition, relation extraction, and event extraction all uniformly adopt a unified set of input-output format constraints, facilitating rapid integration for downstream tasks, stable parsing, and extremely low migration costs.

Reading this, you may ask: aren't general large models already "capable of everything"? Why is a dedicated intent understanding model still needed?

Because in such scenarios, general large models have three "inherent shortcomings."

General large models are accustomed to "thinking it through before answering," often producing lengthy reasoning of hundreds or thousands of words. For a question like "to whom should this ticket be transferred," by the time it finishes speaking, the customer would have long hung up. Xing4.0-29B-Intent responds quickly only with the key decision information needed for task processing, without lengthy reasoning, outputting dozens of tokens on average, with complete result response latency reaching the hundred-millisecond level.

Using a general large model for classification and information extraction is like using a sledgehammer to crack a nut. The intent understanding model is small in scale and economical to invoke; when a scenario requires millions of invocations per day, the cost gap multiplies.

The output of general large models is free text; the same question may yield ten different phrasings, making it difficult for downstream systems to parse stably. The output of the intent understanding model is strictly constrained within a predefined structure: fields are fixed, formats are verifiable, leaving no room for "free improvisation."

In one sentence: where response timeliness requirements are high, scenarios are relatively convergent, and decision execution is relatively clear, the intent understanding model is a faster, cheaper, and more reliable choice than general large models.

There are already pioneers in this track. After the lightweight decision model represented by Jev was launched, it quickly became a focal point of discussion in the developer community. The market's response confirmed a judgment: intent understanding does not need to rely on the "strong reasoning" of large models; dedicated small models can also be fast and accurate.

Whether a model can understand users' fine-grained multi-intent needs can be tested at two levels: first, both judging the category correctly and extracting the specific content correctly; second, looking only at category judgment, whether it can avoid missing selections and avoid extra selections. Based on this, the technical team conducted two sets of evaluations, examining complete extraction results and category judgment capabilities respectively.

The first set of evaluations sampled 6,000 questions proportionally from 48 Chinese and English public datasets, covering four types of tasks: intent understanding, entity recognition, relation extraction, and event extraction. It examines not only "which category it is assigned to" but also "what specifically was extracted."

Taking the lighting command from earlier as an example, merely determining "control lighting" is not enough; information such as "bedroom," "light," and "warm yellow" must also be extracted. This example illustrates the scoring requirement: correct category but missing or incorrect content cannot count as a completely correct extraction result.

The total score is first calculated separately across the four task types, then averaged with equal weights, so it represents the comprehensive performance across the four task types, not the proportion of correct answers calculated by directly merging the 6,000 questions. For all metrics in the table below, higher scores indicate better performance.

Xing4.0-29B-Intent scored 58 in total, higher than the highest total score of 49 among the 18 mainstream general models on the market participating in this evaluation; its sub-scores across the four task types were also all higher than those of the general models participating in this evaluation.

This result shows that under the tasks and settings of this evaluation, Xing4.0-29B-Intent performs better at simultaneously completing category judgment and content extraction, rather than merely being able to select the correct category.

The second set of evaluations does not assess parameter or specific content extraction, but only compares whether models can find all and accurately select the categories that should be output. The evaluation contains 5,606 questions in total, including 3,940 questions with a single correct category and 1,666 questions with multiple correct categories; this batch also includes 839 negative examples with "no extractable content," used to test whether the model falsely reports categories when there is no target content.

Jev participated in the test using two answering methods: "choose one of N" selects one answer from the candidate categories; "judge each one as true or false" sequentially determines whether each candidate category matches. The table below also retains GPT-6-Luna as a general-model reference; the two answering methods refer only to Jev and do not apply to all models in the table.

What really matters in this table is two layers: the overall decision score first compresses the three sub-items of single-choice, multi-choice, and negative examples into one overall decision score, which looks at the model's comprehensive level of "finding the right one, finding all of them, and not falsely reporting"; the three sub-items each then look at one thing—whether single-choice can fix on the sole target, whether multi-choice can select all required categories without missing any, and whether negative examples can hold back and not falsely report when there is no target content. The overall score looks at the whole, the sub-items look at weaknesses, and only when the two align does it show that the model's capability is solid.

On the category decisions in this evaluation, the Xingchen model leads the two comparison models (Jev's two answering methods and GPT-6-Luna) across the board on the overall decision score and the three sub-items of single-choice, multi-choice, and negative examples. The core issue is that the two most error-prone scenarios—multi-target recognition and false-positive prevention on negative examples—are exactly where the Xingchen model pulls ahead: in multi-choice tasks it can more completely lock onto all required categories, and in negative-example scenarios it dares to "hand in a blank sheet," preferring to come away empty-handed rather than force a selection to make up the numbers. This shows that it not only knows "what the user wants to do," but can also make steadier judgments when category boundaries are blurred or information is missing, rather than relying solely on a high hit rate.

*The above evaluation was carried out by the company's technical team. The evaluation results reflect the performance of the participating models under this evaluation's data and test settings, and do not represent their effects on all tasks and actual deployment scenarios. For details on the evaluation experimental environment and test method settings in this article, you can contact the technical team via a message in the backend. Developers and researchers are welcome to exchange views on evaluation methods, result reproduction, and application adaptation.

In terms of specific applications, in the telecom ten-thousand-line customer service system, Xing4.0-29B-Intent undertakes dialogue intent recognition and intelligent ticket routing, providing support for the classification and routing of different customer demands. On terminal products such as smart home speakers, Xing4.0-29B-Intent converts users' spoken commands into structured commands that devices can further process, connecting natural language interaction with device operation.

In the business cold-start stage there is not enough labeled data, and you do not want to label massive samples for each new intent;

Agent / Function Calling applications

need a stable, low-hallucination "semantic entry layer."

Then you are very well suited to put Xing4.0-29B-Intent into your tech stack.

For agent applications that need to call tools, the model can provide intent recognition and function-call parameter extraction capabilities, providing input for subsequent execution.

Xing4.0-29B-Intent is fine-tuned based on the Xingchen general base model Xing4.0-29B-A4B, using specialized corpora to train structured output capabilities. This open-source release includes the following four types of resources:

01 2.753 million supervised fine-tuning training corpus entries

Reconstructed from 61 public datasets, covering 19 domains and 52 languages.

02 Terminal intent understanding corpus

The Xingchen General Artificial Intelligence Laboratory, oriented toward mobile-phone intelligent interaction scenarios, built the Chinese spoken language understanding benchmark CMPA-SLU, and the related results were accepted by EMNLP 2026 as "CMPA-SLU: A Benchmark for Chinese Mobile Personal Assistant Spoken Language Understanding."

Developers are welcome to conduct evaluations and trials around their own scenarios and to provide feedback and suggestions through the code repository.

Code repository:

Code license:

Apache 2.0 The dataset and model weights will be simultaneously opened on Hugging Face

Open-source community access address:

This article is provided by Huawei, and QbitAI is authorized to repost it; the views belong to the original author.