Both a "multiple-choice taker" and a "fill-in-the-blank taker" — Star intention understanding model Xing4.0-29B-Intent released

The Star General Artificial Intelligence Laboratory of China Telecom AI Technology Co., Ltd. officially released and open-sourced the intention understanding specialized model Xing4.0-29B-Intent, simultaneously releasing 2.75 million training corpus entries, model weights, and an inference image. The model combines intention recognition with parameter extraction, converting a user's single sentence into structured information that business systems can continue processing.

Currently, the capabilities of the Xing4.0-29B-Intent model have begun to be applied in terminal products such as the Telecom 10000 customer service hotline and smart home speakers, serving dialogue intention recognition, intelligent work order routing, and spoken command parsing. This open-sourcing will provide more developers with the model and accompanying resources, supporting integration into their own business scenarios.

01 Multiple-choice or fill-in-the-blank

"Help me turn the bedroom light to warm yellow." A single everyday device command contains both what the user wants to do and the specific information needed to complete the operation. The system must extract this information from the natural expression before it can be handed over to subsequent applications for processing.

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

Jev-type decision models mainly "take multiple-choice tests," returning typed decisions by picking answers from given candidates; parameter extraction is akin to "taking fill-in-the-blank tests," extracting the specific information needed to complete the task. Xing4.0-29B-Intent supports both types of tasks, and also supports cases where a single sentence contains multiple intentions, multiple parameters, and multiple entities.

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

02 Why not just use a general-purpose large model?

At this point you might ask: aren't general-purpose large models already "omnipotent"? Why do we need a specialized intention understanding model?

Because in this type of scenario, general-purpose large models have three "inherent shortcomings."

First, output speed

General-purpose large models are used to "thinking things through before answering," often producing long passages of reasoning of hundreds or thousands of words. For a question like "who should this work order be transferred to," by the time it finishes speaking, the customer would have long since hung up. Xing4.0-29B-Intent responds quickly only with the key decision information needed for task processing, without long-form reasoning, outputting an average of a few dozen tokens, with full result response latency achievable at the hundreds-of-milliseconds level.

Second, decision cost

Using a general-purpose large model for classification and information extraction is like using a sledgehammer to crack a nut. Intention understanding models are small in scale and cheap to invoke; when a scenario requires millions of invocations per day, the cost gap multiplies dramatically.

Third, decision reliability

The output of general-purpose large models is free text; the same question may yield ten different phrasings, making stable parsing difficult for downstream systems. The output of intention understanding models is strictly constrained within predefined structures: fields are fixed and formats are verifiable, leaving no room for "free improvisation."

In one sentence: in places with high responsiveness requirements, relatively converged scenarios, and relatively clear decision execution, an intention understanding model is a faster, cheaper, and more reliable choice than a general-purpose large model.

This track already has pioneers. After the launch of lightweight decision models represented by Jev, they quickly became a hot topic of discussion in developer communities. The market's response confirmed a judgment: intention understanding does not need to rely on the "strong reasoning" of large models; specialized small models can also be both fast and accurate.

03 Let the results speak Two sets of evaluations to test recognition and extraction capabilities

Whether a model can understand users' fine-grained, multi-intention needs can be examined 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 omissions and over-selections. The technical team accordingly conducted two sets of evaluations, examining complete extraction results and category judgment ability respectively.

First set: not only must the category be right, the specific content must be right too

The first set of evaluation drew 6,000 questions proportionally from 48 public Chinese and English datasets, covering four types of tasks: intention understanding, entity recognition, relation extraction, and event extraction. It examines not only "which category was assigned" but also "what exactly was extracted."

Taking the earlier lighting command as an example, merely recognizing "light control" is not enough; one must also extract information such as "bedroom," "light," and "warm yellow." This example illustrates the scoring requirements: a correct category with missing or erroneous content does not count as a complete and correct extraction result.

The total score is first calculated separately for the four types of tasks, then averaged with equal weights; it therefore represents the comprehensive performance across the four types of tasks, not the proportion of correct answers computed by directly pooling the 6,000 questions. In the table below, a higher score for each metric indicates better performance.

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

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

Group 2: Category judgment only Whether nothing is missed and nothing is over-selected

The second group of evaluations does not assess parameters or specific content extraction; it only compares whether models can find all and accurately select the categories that should be output. The evaluation comprises 5606 questions, including 3940 questions with a single correct category and 1666 questions with multiple correct categories; among these questions there are also 839 negative-example questions with “no extractable content,” used to test whether models falsely report categories when there is no target content.

Jev participates in the test with two answering methods: “choose 1 out of N” selects one answer from the candidate categories, while “judge each one true or false” assesses each candidate category one by one for whether it matches. The table below also retains GPT-6-Luna as a general-purpose model reference; the two answering methods apply only to Jev and not to all models in the table.

What really needs to be looked at in this table is two layers: the overall decision score first compresses the three sub-items—single choice, multiple choice, and negative examples—into one total decision score, looking at the model’s comprehensive level of “accurate targeting, complete coverage, and no false positives”; then each of the three sub-items looks at one thing—whether single choice can lock onto the sole target, whether multiple choice can tick every category that should be selected without missing any, and whether negative examples can hold back and not falsely report when there is no target content. The total score shows the overall picture, while the sub-items reveal weaknesses; only when the two correspond does it show that the model’s capabilities are solid.

In the category decision-making part of this evaluation, the Xingchen model led comprehensively over the two comparison models (Jev's two answering methods and GPT-6-Luna) in total decision score as well as the three sub-items of single-choice, multi-choice, and negative examples. The core issue is: multi-target recognition and false-positive prevention for negative examples, the two most error-prone scenarios, are precisely where the Xingchen model opens up the gap — in multi-choice tasks it can more completely lock in all categories that should be selected, and in negative-example scenarios it dares to 'hand in a blank sheet', preferring to return empty-handed rather than forcing a selection to make up the numbers. This shows it not only knows 'what the user wants to do', but can also make steadier judgments when category boundaries are ambiguous or information is missing, rather than simply relying on a high hit rate.

*The above evaluation was conducted by our 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 results in all tasks and actual deployment scenarios. For details on the evaluation experiment 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 engage in discussions on evaluation methods, result reproduction, and application adaptation.

04 About applications: Who is this model for?

In terms of specific applications, in the Telecom Wanhao customer service system, Xing4.0-29B-Intent handles dialogue intent recognition and intelligent ticket routing, supporting the classification and routing of different customer requests. On terminal products such as smart home speakers, Xing4.0-29B-Intent converts users' spoken instructions into structured instructions that devices can further process, bridging natural language interaction and device operation.

If you are working on:

Terminal interaction

Phone voice assistants / smart cockpits / IoT control centers

You need both intent judgment and parameter extraction, and you find large models too slow and too expensive;

ToB customer service / enterprise IM bots

In the cold-start stage of your business you lack sufficient labeled data and don't want to label massive samples for every new intent;

Agent / Function Calling applications

You need a stable, low-hallucination 'semantic entry layer'.

Then, Xing4.0-29B-Intent is a perfect fit for your tech stack.

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

05 Open Schema + 2.75 million corpus entries,Welcome to try it out

Xing4.0-29B-Intent is fine-tuned from the XingChen general-purpose base model Xing4.0-29B-A4B, trained with specialized corpora for structured output capability. This open-source release includes the following four types of resources:

01 2,753,000 supervised fine-tuning training samples

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

02 On-device intent understanding corpus

The XingChen AGI Lab built the Chinese spoken language understanding benchmark CMPA-SLU for smart interaction scenarios on mobile devices, and the related work was accepted by EMNLP 2026 as "CMPA-SLU: A Benchmark for Chinese Mobile Personal Assistant Spoken Language Understanding".

03 Xing4.0-29B-Intent model weights

Supporting intent recognition, parameter extraction, and structured output in a unified format.

04 Official inference image

Provided along with the code repository, offering supporting resources for model deployment and trial use.

Developers are welcome to conduct evaluations and trials around their own scenarios, and to give feedback and suggestions via the code repository.

GET STARTED

Tianyi AI Open Platform invocation address:

https://www.teleai.com.cn/doc/1MBaE2xId9JYYxLXUz2R/3jUbsGzfAj0Qe4A38Y0s

Code repository:

https://github.com/XingChen-AGI/Xing4.0-29B-Intent

Code license:

Apache 2.0 The datasets and model weights will also be released on Hugging Face

Open-source community invocation addresses:

HuggingFace: https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent

HuggingFace dataset: https://huggingface.co/datasets/XingChen-AGI/Xing4.0-29B-Intent-oss

GitHub: https://github.com/XingChen-AGI/Xing4.0-29B-Intent

ModelScope: https://www.modelscope.cn/models/XingChen-AGI/Xing4.0-29B-Intent

ModelScope dataset: https://www.modelscope.cn/datasets/XingChen-AGI/Xing4.0-29B-Intent-oss

Modelers: https://modelers.cn/models/XingChen-AGI/Xing4.0-29B-Intent

This article was provided by Huawei and reprinted by QbitAI with authorization; the views belong to the original author.