China Telecom open-sources Xing4.0-29B-Intent, a 29B intent model, with 2.753M-sample training corpus

China Telecom's Xingchen General AI Lab has released and open-sourced Xing4.0-29B-Intent, a 29B-parameter model built for intent recognition and parameter extraction, along with model weights, a 2.753 million-sample training corpus and an official inference image. The source is a Chinese-language article published by QbitAI (量子位) on 2026-10-22. It was supplied by Huawei and reprinted by QbitAI with authorization. The lab belongs to China Telecom Artificial Intelligence Technology Co., Ltd. (中国电信人工智能科技有限公司).

What the model does

Xing4.0-29B-Intent converts a single utterance into structured information a business system can process. For "set the bedroom light to warm yellow," it returns the intent (light control) and the parameters: room (bedroom), device (light) and state (warm yellow). It handles both typed decisions chosen from candidates and parameter extraction, including sentences with multiple intents, parameters and entities.

The model uses a Unified Schema-driven Semantic Parsing (USP) framework. Results are organized as category plus parameters and output as JSON under predefined field rules. Intent understanding, entity recognition, relation extraction and event extraction share one input/output format.

It is fine-tuned from the general base model Xing4.0-29B-A4B. According to the company, average output is a few dozen tokens, with full-result latency in the hundred-millisecond range. Its capability table lists inference latency under 200ms. Context window and API pricing are not stated. Pricing for the Teleai open platform is not yet disclosed.

Benchmarks (company-run)

The company's technical team ran both evaluations and says results reflect only the tested data and settings.

Full extraction test: 6,000 questions sampled proportionally from 48 Chinese and English public datasets, covering intent understanding, entity recognition, relation extraction and event extraction. The total is the equal-weighted average of the four task scores.

Model	Total	Intent	Entity	Relation	Event
Xing4.0-29B-Intent	58	81	72	38	40
Gemini-3.6-Flash	49	64	71	29	30
Qwen3.8-Max	46	63	66	26	29
Kimi-K3	45	61	68	25	28
GPT-5.6-Sol	45	60	68	23	28

The company says Xing4.0-29B-Intent also outscored every general-purpose model it tested on all four sub-tasks.

Category decision test: 5,606 questions (3,940 single-answer, 1,666 multi-answer, including 839 negative cases with nothing to extract).

Model	Overall	Single	Multi	Negative
Xing4.0-29B-Intent	70	65	68	70
Jev	56	60	41	35
Jev (yes/no judging)	53	24	64	5
GPT-6-Luna	65	52	63	64

Open-source release

• 2.753 million supervised fine-tuning samples, rebuilt from 61 public datasets across 19 domains and 52 languages
• Terminal intent-understanding data, including the CMPA-SLU Chinese mobile spoken language understanding benchmark, which the source says was accepted at EMNLP 2026
• Model weights
• Official inference image in the code repository

The code license is Apache 2.0. Resources are listed on GitHub, Hugging Face, ModelScope and Modelers, with API access through the Teleai open platform.

Deployment

The company says the model is already used in China Telecom's 10000-number customer service for dialogue intent recognition and ticket routing, and in smart home speakers for parsing spoken commands.

What this means

The model targets a narrow job: fast, schema-constrained output that downstream systems can validate, rather than free-text reasoning. The company argues this makes it faster, cheaper and more reliable than general models in latency-sensitive, well-bounded scenarios. The benchmark results are self-reported on data the company selected, and the full test setup is available only by contacting its team. Independent replication will determine how far the lead holds. The open corpus, weights and Apache 2.0 code license let developers test it on their own intent schemas.