Xing4.0-29B-Intent and LingDT-Decision-1.0: Why Specialized Decision Models Are Replacing General LLMs for Agent Routing

On October 22, 2026, two major Chinese AI labs released dedicated models built specifically for intent routing and tool invocation: China Telecom's TeleAI open-sourced Xing4.0-29B-Intent under the Apache 2.0 license alongside 2.753 million supervised training samples, while Ant Digital Technologies introduced the LingDT-Decision-1.0 series (4B and 27B), slashing token usage by 90–97% compared to general-purpose chat models on high-frequency enterprise routing tasks.

If you have ever wired up a voice assistant or a multi-agent workflow in production, you know the most frustrating bottleneck. A user says something simple like, "Set the bedroom light to warm yellow and wake me up at seven," and your pipeline forwards that raw string to a massive reasoning LLM. The model dutifully spins up a chain of thought, spends 1.5 seconds pondering lighting ergonomics, burns eight hundred tokens, and finally spits out a JSON payload to trigger two API calls. All you actually needed at the front door was a 150-millisecond classifier that picks the right category and extracts three parameters.

That exact architectural gap is why engineers are splitting agent pipelines in two—separating open-ended generation from compact Decision Models and Intent Extraction Models.

General Chat Models vs. Strict Schemas: Where the Friction Lies

When you are running a side project with twenty requests a day, calling a frontier model to pick a tool is fine. At scale—say, a telecom contact center or a fintech risk engine handling millions of calls daily—general-purpose LLMs hit three stubborn walls:

• First-action latency. Reasoning models are trained to "think before speaking." By the time a large model finishes its internal monologue just to route a support ticket, a caller on a phone line has already hung up.
• Fragile formatting. Free-form text generation occasionally drifts out of schema—missing a bracket or inventing a parameter name your backend parser rejects.
• Wasted compute. Paying frontier token rates to pick one option out of ten or fill three slots in a form drains budgets fast.

We see the exact same preference for fast, lightweight models in anonymized production metrics from NeuralSpace Chat. According to our internal chat_history table over a 30-day window (September 22 to October 21, 2026, sample size N = 852 completed assistant replies), users rely heavily on compact models for everyday turns: GPT-5.4 Mini accounted for 242 replies (averaging 613 tokens at a mean cost of 1.62 tokens per reply), and DeepSeek V4.1 Flash handled 106 replies (averaging 1,170 tokens at just 0.30 tokens per reply). By contrast, the heavy flagship Claude Opus 5 was invoked 31 times for deep, long-form tasks (averaging 2,278 tokens and 31.20 tokens per reply). A similar pattern holds in Vibe Coding: across 54 anonymized vibe_chats sessions over 90 days (July 23 to October 21, 2026), fast models led adoption with DeepSeek V4 Flash (21 sessions) and DeepSeek V4.1 Flash (10 sessions).

Inside Xing4.0-29B-Intent: Multiple Choice Meets Fill-in-the-Blank

Built by TeleAI's XingChen lab on top of the Xing4.0-29B-A4B Mixture-of-Experts backbone (29 billion total parameters with 4 billion active per token), Xing4.0-29B-Intent introduces a framework called USP (Unified Schema-driven Semantic Parsing).

Instead of chaining separate microservices for intent classification, named entity recognition (NER), relation extraction, and function-call slot filling, USP folds all of them into a single pass. You pass a schema definition, and the model returns a strictly validated category + args JSON object in under 200 milliseconds—generating only a few dozen tokens without verbose reasoning traces.

The team evaluated the model across two benchmark suites:

1. Joint classification and extraction: 6,000 test cases sampled across 48 public bilingual datasets covering intent, NER, relation, and event extraction. Getting the category right while missing a single parameter counts as a failure. Xing4.0-29B-Intent scored 58 overall, beating 18 general-purpose models including Gemini 3.6 Flash (49), Qwen 3.8 Max (46), Kimi K3 (45), and GPT-5.6 Sol (45).
2. Multi-label and negative-sample discipline: 5,606 questions including 1,666 multi-intent queries and 839 negative traps where no valid category applies. Here the model reached a decision score of 70, ahead of GPT-6 Luna (65) and the lightweight decision router Jev (56). Its biggest edge came on negative examples (70 vs. Jev's 35)—knowing when to return an empty result instead of forcing a false positive.

Alongside the Apache 2.0 model weights and inference container, TeleAI released the Xing4.0-29B-Intent-oss dataset on Hugging Face and GitHub: 2.753 million supervised fine-tuning samples restructured from 61 public datasets across 19 domains and 52 languages, plus the CMPA-SLU spoken-language benchmark accepted at EMNLP 2026.

LingDT-Decision-1.0: Native Multi-Choice and DAG Execution

Released the same day by Ant Digital Technologies in 4B and 27B sizes, LingDT-Decision-1.0 targets compound business requests in customer service, risk control, and finance.

Consider a real support message: "I want a refund, cancel auto-renewal, and reissue my invoice, but don't delete my account." A single-label classifier tags the whole ticket as "Refund" and drops the rest. Worse, those actions have strict dependencies: you must verify the order before issuing a refund, and the invoice depends on the refund outcome, whereas canceling auto-renewal can run right away.

LingDT-Decision-1.0 solves this with native multi-choice selection and ordered decision-making via a Directed Acyclic Graph (DAG). Across 23 open-source benchmarks, the 27B version ranked first overall ahead of Microsoft-Decision-1, Jev 1.13.0, and Fastino/GLiDE. In production tests reported by Ant Digital:

• data labeling pipelines saw a 97% drop in token consumption and a 10x speedup compared to DS-V4-Flash;
• customer support workflows cut tokens by 90%, doubled processing speed, and lifted risk-detection accuracy from 76% to 89% against Qwen-3.7-Flash;
• financial tool-calling tasks saw a 5x speedup, a 2.6x increase in exact tool-chain match rate, and hallucination recall jumping from 44% to 70%.

Where Decision Models Fall Short—and How to Use the Pattern

To be fair, dedicated decision models have a hard ceiling. They cannot write marketing copy, brainstorm creative video concepts, or handle open-ended troubleshooting when a user's problem falls outside predefined schemas. If your business rules change hourly without updating candidate definitions, a strict schema router will simply reject valid requests.

Where they shine is acting as the fast semantic gatekeeper in front of heavier models. Whether you are building pipelines through the NeuralSpace API or orchestrating custom AI Agents, the takeaway is practical: let compact Flash and Mini models handle intent parsing and JSON slot extraction in milliseconds, and save your frontier reasoning models for the steps that genuinely require deep synthesis.

Primary sources: QbitAI report on Xing4.0-29B-Intent, XingChen-AGI GitHub repository, and InfoQ coverage of LingDT-Decision-1.0.