Skip to content
Main Site News Console

Combining "Selection" and "Fill-in-the-Blank" Capabilities: Xingchen Intent Understanding Model Xing4.0-29B-Intent Officially Released

· 量子位
国内AI

China Telecom AI Releases and Open-Sources Xing4.0-29B-Intent: A Dedicated Intent Understanding Model

Image 3

The Xingchen AGI Lab at China Telecom Artificial Intelligence Technology Co., Ltd. has officially released and open-sourced Xing4.0-29B-Intent, a dedicated intent understanding model, alongside 2.75 million training samples, model weights, and inference container images. By deeply integrating intent classification with parameter extraction, the model swiftly converts natural language user instructions into structured data ready for direct downstream execution.

Currently, Xing4.0-29B-Intent has been deployed in production across China Telecom’s 10000 customer service hotline, smart home speakers, and other edge devices—powering conversational intent classification, intelligent ticket routing, and spoken instruction parsing. This open-source release provides developers with the complete model and supporting resources to accelerate integration into their own production environments.


01 True-or-False or Fill-in-the-Blank?

“Turn the bedroom light to warm yellow.”

A routine device control command contains both the action the user intends to execute and the specific parameters required to complete it. The system must accurately extract this key information from natural language before handing it off to downstream applications for execution.

When processing such a command, merely classifying the intent as “control lighting” is far from sufficient. The system must also accurately extract:

  • Which room? — Bedroom
  • Which device? — Light
  • Which target state? — Warm yellow

Decision models like Jev primarily handle “multiple-choice” tasks—returning categorical decisions by selecting from a predefined list of candidates. Parameter extraction, on the other hand, is like “filling in the blanks”—extracting the exact parameters required to perform the task. Xing4.0-29B-Intent natively handles both tasks, robustly resolving complex utterances that feature multiple intents, multiple parameters, and multiple entities.

The model adopts a Unified Schema-driven Semantic Parsing (USP) framework, organizing results into a hierarchical “category + parameter” structure and outputting JSON-formatted data aligned with predefined field schemas. Across heterogeneous objectives—including intent understanding, named entity recognition (NER), relation extraction, and event extraction—the model enforces a unified, standard input-output format, substantially lowering integration and migration costs while guaranteeing parsing stability.


02 Why Not Just Use General-Purpose LLMs?

At this point, you might wonder: If general-purpose LLMs can already do “everything,” why do we need a dedicated intent understanding model?

In real-world enterprise deployments, general-purpose LLMs exhibit three significant inherent bottlenecks:

1. Latency and Generation Speed

General-purpose LLMs tend to overthink, frequently generating reasoning chains hundreds or thousands of tokens long. In latency-critical real-time scenarios—such as rapid ticket dispatching—waiting for an LLM to complete inference might mean the caller has already hung up. Xing4.0-29B-Intent focuses exclusively on distilling the key decision parameters required for the task, eliminating redundant reasoning. With an average output of only a few dozen tokens, end-to-end response times can be pushed down to the hundred-millisecond range.

2. Inference Cost

Using a massive general-purpose LLM for basic classification and information extraction is like using a sledgehammer to crack a nut. Dedicated intent models feature a manageable parameter scale and minimal inference overhead. In production workloads handling millions of requests per day, this cost advantage compounds exponentially.

3. Determinism and Reliability

The unconstrained text generation of general-purpose LLMs is inherently non-deterministic; identical inputs can yield varying phrasing, creating headaches for downstream structured parsing. In contrast, Xing4.0-29B-Intent strictly constrains its output to predefined schemas: fields are well-defined, formats are strictly controlled, and outputs pass rigorous validation, effectively eliminating hallucination risks caused by open-ended generation.

In short: In environments with strict latency budgets, well-defined business domains, and explicit actions, specialized intent understanding models are faster, more cost-effective, and far more reliable than general-purpose LLMs.

Lightweight decision models like Jev have previously sparked widespread community interest. Their market traction validates an emerging industry consensus: intent understanding does not require blind reliance on heavy LLM reasoning—lightweight, dedicated models can be just as fast and accurate.


03 Performance Speaks: Two Benchmark Suites Evaluating Classification and Extraction

Whether a model can accurately understand fine-grained, multi-intent user requests was evaluated across two key dimensions:

  1. Balancing Classification and Extraction: Correctly classifying the intent while precisely extracting the required parameters;
  2. Pure Classification Capability: Accurately predicting categories without missing labels, over-predicting, or generating false positives.

The engineering team designed two specialized benchmark suites targeting these dimensions.

Benchmark 1: Getting Both the Category and the Content Right

The first benchmark sampled 6,000 test cases proportionally from 48 public Chinese and English datasets, spanning four representative tasks: intent understanding, entity detection, relation detection, and event extraction. The evaluation focused on whether the model could correctly identify the category while completely extracting the corresponding entities and parameters.

Taking the lighting command as an example, merely identifying “control lighting” scores zero; it must also extract parameters such as “bedroom”, “light”, and “warm yellow”. If the category is correct but the extracted content is incomplete or incorrect, it is marked as a failure.

Scores were calculated independently within each of the four task types and then averaged with equal weighting (to reflect balanced overall capability rather than a simple unweighted aggregate). The results are detailed below (higher scores indicate better performance):

ModelTotal ScoreIntent ClassificationEntity DetectionRelation DetectionEvent Extraction
Xing4.0-29B-Intent5881723840
Gemini-3.6-Flash4964712930
Qwen3.8-Max4663662629
Kimi-K34561682528
GPT-5.6-Sol4560682328
GLM-5.24458652529
GPT-5.54459672328
GLM-5.3-Flash4458652627
GLM-5.14259602427
DeepSeek-V4.1-Flash4256612326
DeepSeek-V4-Pro4156602423
LongCat-2.04153622324
GPT-4.14055592223
DeepSeek-V4-Flash3952552425
GPT-6-Luna3951592024
GPT-4.1-Mini3749572121
Qwen3.7-Flash3648552220
GPT-4o-Mini3141491816

The benchmark results show that when simultaneously handling categorical classification and fine-grained content extraction, Xing4.0-29B-Intent significantly outperforms mainstream general-purpose models.

Benchmark 2: Pure Classification — Precision and Recall Without Over-Selection

The second benchmark isolated classification capability by removing the parameter extraction component, focusing strictly on the model’s ability to identify all correct categories without false positives. The test set contained 5,606 questions (3,940 single-label and 1,666 multi-label cases) alongside 839 negative samples with “no extractable content” to evaluate false-positive behavior when target entities are absent.

Jev was evaluated using both “1-of-N selection” and “per-category binary classification” strategies, with GPT-6-Luna included as a general-purpose LLM baseline:

ModelTotal Decision ScoreSingle-Choice ScoreMultiple-Choice ScoreNegative Sample Score
Xing4.0-29B-Intent70656870
Jev (1-of-N)56604135
Jev (Binary Classification)5324645
GPT-6-Luna65526364

Test results demonstrate that Xing4.0-29B-Intent leads across the board in total decision score as well as single-choice, multiple-choice, and negative sample sub-metrics. It demonstrates a distinct advantage in multi-target recognition and false-positive suppression: reliably covering all valid categories in multi-label scenarios while decisively refusing to predict when faced with negative samples, effectively avoiding false positives caused by forced matching.

Feature Comparison

CapabilityDecision Models (e.g., Jev)Xing4.0-29B-Intent
Intent ClassificationSupportedSupported
Function Calling Parameter ExtractionNot SupportedSupported
Information Extraction / NER / Event ExtractionNot SupportedSupported
Fixed-Schema Structured OutputSupportedSupported
Inference Latency< 100ms< 200ms

04 Target Applications: Who Is This Model For?

In production at China Telecom, Xing4.0-29B-Intent already handles conversational intent classification and intelligent ticket dispatching for the 10000 customer service hotline. On IoT edge devices such as smart home speakers, it reliably maps natural spoken instructions into low-level device control protocols.

If your team is building:

  • Device Interactions: Smartphone voice assistants, smart cockpits, and IoT central control hubs (requiring simultaneous intent recognition and parameter extraction while remaining highly sensitive to latency and inference costs);
  • Enterprise Customer Service / Corporate IM Bots: Systems in early cold-start phases lacking large-scale labeled samples that cannot absorb the high cost of per-intent annotation;
  • AI Agents / Function Calling Applications: Architectures requiring a robust, low-hallucination, and trusted upstream “semantic gateway” to route and dispatch tool calls.

Then Xing4.0-29B-Intent represents an ideal foundational choice for your tech stack.


05 Open Schema + 2.75M Training Samples: Ready for Exploration

Fine-tuned from the Xingchen general base model Xing4.0-29B-A4B, Xing4.0-29B-Intent features specialized enhancements for intent parsing and structured output generation. This open-source release includes four core assets:

  1. 2.753 Million Supervised Fine-Tuning (SFT) Samples: Reconstructed from 61 public datasets, covering 19 industry verticals and 52 languages.
  2. On-Device Intent Understanding Benchmark: CMPA-SLU, a Chinese spoken language understanding benchmark designed for mobile intelligent interactions (research paper accepted by EMNLP 2026).
  3. Full Model Weights: Native support for intent classification, parameter extraction, and standardized structured JSON output.
  4. Official Deployment Containers: Provided directly in the open-source repository for turnkey inference and production serving.