Xing4.0-29B-Intent

Xing4.0-29B-Intent#

🦉 GitHub • 🤗 Hugging Face • 🤖 ModelScope • 🪐 Modelers • 📚 Dataset

For full project documentation, deployment instructions, SDK usage, and data details, visit the GitHub repository.

Contents#

• Introduction
• Model Information
• Evaluation Results
• Quick Start
• Companion Dataset
• License and Citation

Introduction#

Xing4.0-29B-Intent is a schema-guided model for intent recognition and structured information extraction, developed from Xing4.0-29B-A4B, the Xing semantic large language model from China Telecom Artificial Intelligence Technology Co., Ltd. It converts natural language into structured results for business systems, supporting intelligent assistants, request routing, and document information extraction.

For example, a request such as “Find flights from Beijing to Shanghai tomorrow” needs to be parsed into a flight-search intent and parameters such as origin, destination, and date. Xing4.0-29B-Intent performs recognition and extraction according to the supplied category and field definitions, returning results in a unified JSON format.

Key Features#

• Four task types, one representation: intent recognition and slot filling, named entity recognition, relation extraction, and event extraction all use category + args to represent results.
• Schema-guided task configuration: specify tasks through candidate categories, category descriptions, and field definitions, using the same input/output protocol across business scenarios.
• Joint recognition and extraction: identify user intent and its parameters together, with JSON arrays supporting multiple categories and multiple argument instances.
• Chinese and English use cases: evaluation covers Chinese and English corpora, including intent understanding in conversations and structured information extraction from text.
• Companion SDK: provides schema definitions, prompt assembly, multi-turn conversation input, synchronous and asynchronous calls, and convenient access to structured results.

Model Information#

The base model uses a mixture-of-experts (MoE) architecture. See the base model overview for technical details. This project focuses on intent recognition and information extraction; performance for specific use cases should be validated using the target data and schema.

Tasks and Outputs#

The input schema defines candidate categories and their extractable fields: category specifies the category name, while args lists field names and descriptions. Model outputs use the corresponding category and field names. Each object in the output args array represents one extracted argument instance; when no arguments can be filled, args is [].

Evaluation Results#

We evaluate the model on four task types: intent understanding, named entity recognition, relation extraction, and event extraction. The table below presents the currently compiled results on a 0–100 scale, rounded to whole numbers from the original scores.

Note: Bold marks the highest original score in each column, determined before rounding. Identical displayed integers do not necessarily indicate a tie in the original scores.

Evaluation Notes#

• Evaluation data: the test set contains 6,000 samples drawn from 48 public datasets, covering named entity recognition, relation extraction, event extraction, and intent understanding. It evaluates the model's ability to identify categories and extract relevant information from input text according to the supplied task instructions and schema.
• Metric definitions: each task score is the macro-average of per-dataset field-value Micro-F1 scores. The overall score is the equally weighted average of the four task scores. The category score evaluates only category labels and is also aggregated with equal weights across the four task types. Scores are shown on a 0–100 scale and do not represent business accuracy rates.
• Interpretation: category scores do not imply that all extracted fields are correct. Interfaces, prompts, and inference settings may also affect results.

Quick Start#

Start the Model Service#

For environment setup and service startup, refer to the base model's serving instructions (Chinese), which cover compatible environments and deployment instructions for frameworks such as vLLM and SGLang.

Replace the model path in those instructions with the path to the Xing4.0-29B-Intent weights, and set --served-model-name to Xing4.0-29B-Intent. Follow the base model's deployment documentation for framework compatibility requirements, and configure GPU memory usage, parallelism, and context length for your environment.

Once the service is running, the SDK calls it through its OpenAI-compatible API. The example below uses http://localhost:8000/v1.

Note. Upstream integration status (as of 2026-11-10):

• SGLang — native support for the base architecture merged into the mainline (sgl-project/sglang#39793, merged 2026-11-05); the out-of-tree plugin is still available for earlier versions.
• vLLM — support ships as an out-of-tree plugin: XingChen-AGI/xing4-vllm-plugin. A stock upstream vLLM installation does not register xing4_0 out of the box, so install the plugin into the same Python environment before serving.

Prebuilt container images and step-by-step guides are listed in the GitHub repository.

Install and Use the SDK#

The companion XingIntent SDK lets you define business intents and retrieve structured results. Clone the GitHub repository and install the SDK:

git clone https://github.com/XingChen-AGI/Xing4.0-29B-Intent.git
cd Xing4.0-29B-Intent/XingIntentSDK
python -m pip install -e .

Then call the model service with the following example. Replace base_url and model with your service address and model name. The Chinese input and schema are retained from the Chinese README so that both versions show the same example.

from xingintent import Arg, Intent, IntentClient

client = IntentClient(
    base_url="http://localhost:8000/v1",
    model="Xing4.0-29B-Intent",
    max_tokens=512,
)
result = client.predict(
    text="帮我查明天北京到上海的航班。",
    intents=[
        Intent(
            "查询航班",
            "查询指定日期与路线的航班",
            args=[
                Arg("出发地", "出发城市"),
                Arg("目的地", "到达城市"),
                Arg("日期", "用户提到的出行日期，保留原文"),
            ],
        )
    ],
)
print(result.category)
print(result.args)

Expected printed output:

查询航班
{'出发地': '北京', '目的地': '上海', '日期': '明天'}

result.category and result.args are convenience properties provided by the SDK. They return the first predicted category and its merged argument dictionary, respectively. For multiple categories or multiple argument instances, iterate over result.intents and read each item's category and args to preserve the full structure.

The SDK handles prompt assembly and result parsing automatically. For multi-turn conversations, asynchronous calls, loading schemas from files, and constrained decoding, see the SDK documentation. More examples are available under SDK examples.

Companion Dataset#

The companion open-source dataset, Xing4.0-29B-Intent-oss, contains reconstructed public datasets and China Telecom's in-house domain data. It helps developers understand the task definitions, data organization, and input/output formats.

• Hugging Face dataset page: dataset overview.
• Dataset documentation: data composition, schemas, field definitions, and example assembly instructions.

License and Citation#

License#

The model weights, public data, and code (including the Python SDK) are released under the Apache License 2.0.

Citation#

If this project helps your research or applications, please consider citing it and the related work on its base model.

@misc{li2026xinguspcrossfamilysemantic,
      title={Xing USP: A Cross-Family Semantic Contract for Information Extraction and Intent-Slot Parsing}, 
      author={Zhongqiu Li and Boyang Wang and Ruiyu Fang and Shiquan Wang and Di Wu and Jiayi Liang and Feilong Zhang and Hongfeng Du and Xiaomeng Huang and Mengjiao Bao and Luwen Pu and Xuewei He and Yu Zhao and Chao Wang and Kaidong Yu and Bingkai Yang and Qingjun Zhu and Shaoyi Du and Shuangyong Song and Yongxiang Li},
      year={2026},
      eprint={2610.34343},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.34343}, 
}

@misc{liu2025trainingreporttelechat3moe,
      title={Training Report of TeleChat3-MoE}, 
      author={Xinzhang Liu and Chao Wang and Zhihao Yang and Zhuo Jiang and Xuncheng Zhao and Haoran Wang and Lei Li and Dongdong He and Luobin Liu and Kaizhe Yuan and Han Gao and Zihan Wang and Yitong Yao and Sishi Xiong and Wenmin Deng and Haowei He and Kaidong Yu and Yu Zhao and Ruiyu Fang and Yuhao Jiang and Yingyan Li and Xiaohui Hu and Xi Yu and Jingqi Li and Yanwei Liu and Qingli Li and Xinyu Shi and Junhao Niu and Chengnuo Huang and Yao Xiao and Ruiwen Wang and Fengkai Li and Luwen Pu and Kaipeng Jia and Fubei Yao and Yuyao Huang and Xuewei He and Zhuoru Jiang and Ruiting Song and Rui Xue and Qiyi Xie and Jie Zhang and Zilu Huang and Zhaoxi Zhang and Zhilong Lu and Yanhan Zhang and Yin Zhang and Yanlei Xue and Zhu Yuan and Teng Su and Xin Jiang and Shuangyong Song and Yongxiang Li and Xuelong Li},
      year={2025},
      eprint={2512.24157},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.24157}, 
}

@misc{wang2025technicalreporttelechat2telechat25,
      title={Technical Report of TeleChat2, TeleChat2.5 and T1}, 
      author={Zihan Wang and Xinzhang Liu and Yitong Yao and Chao Wang and Yu Zhao and Zhihao Yang and Wenmin Deng and Kaipeng Jia and Jiaxin Peng and Yuyao Huang and Sishi Xiong and Zhuo Jiang and Kaidong Yu and Xiaohui Hu and Fubei Yao and Ruiyu Fang and Zhuoru Jiang and Ruiting Song and Qiyi Xie and Rui Xue and Xuewei He and Yanlei Xue and Zhu Yuan and Zhaoxi Zhang and Zilu Huang and Shiquan Wang and Xin Wang and Hanming Wu and Mingyuan Wang and Xufeng Zhan and Yuhan Sun and Zhaohu Xing and Yuhao Jiang and Bingkai Yang and Shuangyong Song and Yongxiang Li and Zhongjiang He and Xuelong Li},
      year={2025},
      eprint={2507.18013},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18013}, 
}

@misc{he2024telechattechnicalreport,
      title={TeleChat Technical Report}, 
      author={Zhongjiang He and Zihan Wang and Xinzhang Liu and Shixuan Liu and Yitong Yao and Yuyao Huang and Xuelong Li and Yongxiang Li and Zhonghao Che and Zhaoxi Zhang and Yan Wang and Xin Wang and Luwen Pu and Huinan Xu and Ruiyu Fang and Yu Zhao and Jie Zhang and Xiaomeng Huang and Zhilong Lu and Jiaxin Peng and Wenjun Zheng and Shiquan Wang and Bingkai Yang and Xuewei he and Zhuoru Jiang and Qiyi Xie and Yanhan Zhang and Zhongqiu Li and Lingling Shi and Weiwei Fu and Yin Zhang and Zilu Huang and Sishi Xiong and Yuxiang Zhang and Chao Wang and Shuangyong Song},
      year={2024},
      eprint={2401.03804},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2401.03804}, 
}