Back 量子位 AI 报道

量子位 AI 报道· 2026-10-22 15:11 · Original publication time

XingChen General AI Lab releases Xing4.0-29B-Intent with open model weightsMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI-assisted summary

The model extracts intents and parameters from a request into structured information for uses including customer service ticket routing and smart home commands. The lab has also made model weights, an inference image and about 2.75 million training records available. Its published evaluation was conducted by the company’s own team and does not represent performance in every deployment setting.

Article · AI translation

China Telecom AI Technology Co., Ltd.'s Xingchen General Artificial Intelligence Laboratory officially releases and open-sources the specialized intent understanding model Xing4.0-29B-Intent

China Telecom AI Technology Co., Ltd.'s Xingchen General Artificial Intelligence Laboratory has officially released and open-sourced the specialized intent understanding model Xing4.0-29B-Intent, while making available 2.75 million training examples, model weights, and an inference image. The model combines intent recognition with parameter extraction, turning a user's sentence into structured information that a business system can process further.

Xing4.0-29B-Intent's capabilities have begun to be used in products including China Telecom's 10000 customer service system and smart home speakers, for conversational intent recognition, intelligent ticket routing, and spoken command parsing. This open-source release will provide more developers with the model and supporting resources for integration into their own business scenarios.

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

“Help me set the bedroom light to warm yellow.” An everyday device command contains both what the user wants to do and the specific information needed to do it. The system needs to extract this information from natural language before passing it to subsequent applications.

For this sentence, merely identifying “control the lights” is not enough. The system also needs to know: Which room? The bedroom. What device? The light. Set it to what state? Warm yellow.

Decision models such as Jev mainly “answer multiple-choice questions,” returning typed decisions by selecting answers from given candidates; parameter extraction is like “answering fill-in-the-blank questions,” extracting the specific information needed to complete a task. Xing4.0-29B-Intent supports both types of tasks, as well as sentences containing multiple intents, multiple parameters, and multiple entities.

The model uses a Unified Schema-driven Semantic Parsing framework (USP), organizing results in a hierarchical “category + parameters” structure and outputting information in JSON format according to predefined field rules. Different task objectives, such as intent understanding, entity recognition, relation extraction, and event extraction, all use one unified set of input and output format constraints, making it easy for downstream tasks to integrate quickly, parse reliably, and migrate at very low cost.

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

At this point, you may ask: Aren't general-purpose large models already “capable of everything”? Why is a dedicated intent understanding model still needed?

Because general-purpose large models have three “inherent shortcomings” in these scenarios.

General-purpose large models tend to “think things through before answering,” often producing hundreds or thousands of words of lengthy reasoning. For a question such as “Who should receive this ticket?”, the customer would have hung up long before the model finished. Xing4.0-29B-Intent quickly returns only the key decision information needed to handle the task, without lengthy reasoning. Its average output is a few dozen tokens, and it can return a complete result with response latency on the order of hundreds of milliseconds.

Using a general-purpose large model for classification and information extraction is like using a sledgehammer to crack a nut. Intent understanding models are smaller and cheaper to call; when a scenario requires over a million calls a day, the cost difference grows manyfold.

General-purpose large models produce free-form text, and the same question may get ten different phrasings, making reliable parsing difficult for downstream systems. An intent understanding model's output is strictly constrained to a predefined structure: its fields are fixed, its format can be validated, and there is no room for “improvisation.”

In short: Where timely responses are required, scenarios are relatively well defined, and decision execution is relatively clear, an intent understanding model is a faster, cheaper, and more reliable choice than a general-purpose large model.

This field already has pioneers. After lightweight decision models represented by Jev were launched, they quickly became a hot topic in the developer community. The market response supports one conclusion: intent understanding need not rely on the “strong reasoning” of large models; dedicated small models can be both fast and accurate.

03 Let the results speak: Two evaluations test recognition and extraction capabilities

Whether a model can understand a user's fine-grained needs involving multiple intents can be tested on two levels: first, whether it identifies the right categories and extracts the right specific content; second, considering category decisions alone, whether it avoids both omissions and extra selections. The technical team therefore conducted two evaluations, examining complete extraction results and category decision capabilities separately.

The first evaluation proportionally sampled 6000 questions from 48 public Chinese- and English-language datasets, covering four tasks: intent understanding, entity recognition, relation extraction, and event extraction. It examined not only “which category was assigned” but also “what specific content was extracted.”

Take the earlier lighting command: identifying only “control the lights” is insufficient; information such as “bedroom,” “light,” and “warm yellow” must also be extracted. This example illustrates the scoring requirement: if the category is correct but the content is missing or wrong, the extraction result cannot be counted as fully correct.

Scores are calculated separately for the four task types and then averaged with equal weight. The total score therefore represents combined performance across the four tasks, rather than the proportion correct across all 6000 questions pooled together. For every metric in the table below, a higher score indicates better performance.

Xing4.0-29B-Intent scored 58 overall, above the highest overall score of 49 among the 18 mainstream general-purpose models evaluated this time. Its scores on all four task types were also higher than those of the general-purpose models evaluated this time.

Model

Overall score

Intent classification

Entity detection

Relation detection

Event extraction

Model

Overall score

Intent classification

Entity detection

Relation detection

Event extraction

Xing4.0-29B-Intent

Gemini-3.6-Flash

Qwen3.8-Max

Kimi-K3

GPT-5.6-Sol

GLM-5.2

GPT-5.5

GLM-5.3-Flash

GLM-5.1

DeepSeek-V4.1-Flash

DeepSeek-V4-Pro

LongCat-2.0

GPT-4.1

DeepSeek-V4-Flash

GPT-6-Luna

GPT-4.1-Mini

Qwen3.7-Flash

GPT-4o-Mini

This result indicates that, under the tasks and settings of this evaluation, Xing4.0-29B-Intent performed better at both category selection and content extraction, rather than merely selecting the right category.

The second evaluation did not assess parameter or specific content extraction; it compared only whether models could find all the categories that should be output and select them accurately. It comprised 5606 questions, including 3940 with one correct category and 1666 with multiple correct categories. These questions also included 839 negative examples with “no extractable content,” used to test whether models would falsely report a category when no target content was present.

Jev participated in the test using two answer methods: “choose 1 from N,” selecting one answer from candidate categories, and “judge each as yes or no,” determining in turn whether each candidate category applies. The table below also includes GPT-6-Luna as a general-purpose model reference. The two answer methods refer only to Jev and do not apply to every model in the table.

There are two levels to examine in this table: the overall decision score first combines the single-choice, multiple-choice, and negative-example subscores into one score, measuring how well a model “selects accurately, finds all applicable categories, and avoids false positives.” Each subscore then examines one thing: whether single-choice questions identify the sole target, whether multiple-choice questions select every applicable category without omission, and whether negative examples avoid false positives when there is no target content. The overall score shows aggregate performance, while the subscores reveal weaknesses; alignment between the two indicates solid model capability.

Model

Overall decision score

Single-choice score

Multiple-choice score

Negative-example score

Model

Overall decision score

Single-choice score

Multiple-choice score

Negative-example score

Xing4.0-29B-Intent

Jev

Jev (yes/no judgment)

GPT-6-Luna

In category decisions in this evaluation, the Xingchen model outperformed both comparison models (Jev with its two answer methods and GPT-6-Luna) on the overall decision score and all three subscores: single-choice, multiple-choice, and negative examples. The key is that the Xingchen model widened its lead in two particularly error-prone scenarios: identifying multiple targets and avoiding false positives on negative examples. It more completely identified all applicable categories in multiple-choice tasks and was willing to “submit a blank answer” on negative examples, choosing no category rather than forcing a selection. This shows that it not only knows “what the user wants to do” but also makes more reliable judgments when category boundaries are unclear or information is missing, rather than relying solely on a high hit rate.

Capability

Decision models such as Jev

Xing4.0-29B-Intent

Capability

Decision models such as Jev

Xing4.0-29B-Intent

Intent classification

Supported

Intent classification

Supported

Function Call parameter extraction

Not supported

Supported

Function Call parameter extraction

Not supported

Supported

Information extraction / NER / event extraction

Not supported

Supported

Information extraction / NER / event extraction

Not supported

Supported

Fixed-format structured output

Supported

Fixed-format structured output

Supported

Inference latency

<100ms

<200ms

Inference latency

<100ms

<200ms

*The above evaluations were conducted by our company's technical team. The results reflect the performance of the participating models on the data and under the test settings of this evaluation; they do not represent performance across all tasks or real-world deployment scenarios. For details of the experimental environment and test methods used in this article, leave a message through the back-end channel to contact the technical team. Developers and researchers are welcome to discuss evaluation methods, result reproduction, and application adaptation.

04 Applications: Who is this model for?

In practical applications, Xing4.0-29B-Intent handles conversational intent recognition and intelligent ticket routing in China Telecom's 10000 customer service system, supporting the classification and routing of different customer requests. In terminal products such as smart home speakers, Xing4.0-29B-Intent converts users' spoken commands into structured instructions that devices can process further, linking natural-language interaction with device operation.

If you are working on:

Terminal interaction

Mobile voice assistants / smart vehicle cockpits / IoT control hubs

You need to identify intents and extract parameters, and find large models too slow and too expensive;

B2B customer service / enterprise IM bots

You do not have enough labeled data during the initial launch of a business application and do not want to label vast numbers of examples for every new intent;

Agent / Function Calling applications

You need a reliable “semantic entry layer” with few hallucinations.

Then Xing4.0-29B-Intent is an excellent fit for your technology stack.

For agent applications that need to call tools, the model can identify intents and extract function-call parameters, providing input for subsequent execution.

05 Open Schema + 2.75 million data samples; give it a try

Xing4.0-29B-Intent is fine-tuned from the Xingchen general-purpose base model Xing4.0-29B-A4B and uses dedicated training data to develop structured-output capabilities. This open-source release includes the following 4 types of resources:

01 2.753 million supervised fine-tuning training samples

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

02 Terminal intent-understanding data

For intelligent interactions on mobile phones, the Xingchen General Artificial Intelligence Laboratory developed CMPA-SLU, a benchmark for understanding spoken Chinese. The related paper, “CMPA-SLU: A Benchmark for Chinese Mobile Personal Assistant Spoken Language Understanding,” was accepted by EMNLP 2026.

03 Xing4.0-29B-Intent model weights

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

04 Official inference image

Provided with the code repository as a supporting resource for deploying and trying the model.

Developers are welcome to evaluate and try it in their own use cases, and to submit issues and suggestions through the code repository.

GET STARTED

Tianyi AI Open Platform access URL:

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 dataset and model weights will also be made available on Hugging Face.

Open-source community access URLs:

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 republished by QbitAI with authorization. The views expressed are those of the original author.

Found an error? Send a correction