Ant International replaces a multi-step merchant support pipeline with one fine-tuned Qwen3 model

The problem

Ant International, a global payments and financial company, runs a merchant service agent that handles inquiries about cross-border payments, refund disputes and e-commerce settlements. The agent was built as a chain of separate modules: route the question, retrieve information, plan, act, write the reply and review it. The authors say this kind of chain collects patches for each odd case, makes errors hard to trace and adds delay at every step. A mistake early in the chain flows into later steps, and the delay frustrates customers who are worried about money.

What they built

The team replaced the chain with one model, which it calls OneModel. A single model handles the whole exchange: understanding the request, using tools, following company procedures and writing the reply. Business rules and standard operating procedures are built into the model's parameters through further training, instead of sitting in separate modules.

The model is an 8-billion-parameter model based on Alibaba's open-weight Qwen3 family. It is served on Ant International's own infrastructure using NVIDIA H200 GPUs and the open-source vLLM software.

How they did it

Training had three phases. The first taught the model the company's stable domain knowledge using manuals, internal FAQs and past problem cases. The second used about 200,000 examples drawn from top-quality service logs and procedures to teach the model to apply business procedures step by step. The third used reinforcement learning, where a simulated customer talks to the model and a scoring system rewards good behavior, to fix hard edge cases and make replies sound less robotic.

Fast-changing data such as exchange rates is not trained into the model. It is supplied at the moment of each request, which avoids retraining whenever the numbers change.

Results

Ant ran an online A/B test over multiple weeks, comparing OneModel with the existing modular pipeline on the Global Merchant Service Agent. Latency fell by more than 50%, from 18.7s to 8s, and the Intelligent Resolution Rate rose from 64.3% to 83.3%. The paper also reports that the 8B OneModel reached a peak resolution rate of 90.75% in its evaluation, above GPT-5.2 at 87.55%, in an evaluation whose setup the paper does not describe.

What went wrong

The authors list lessons from the move. A simulated customer that sounded human (95.8% human-likeness, against 96.7% for real people) did not behave like a real one. Real customers asked for a human agent about 10% of the time and the simulator only 1.7%, so the team had to train it to show impatience and complaints. Noisy human labels also hurt the scoring model used in training, so the team built a loop that cleans the labels. Finally, one model is harder to debug than a chain of modules, because it is no longer clear which step failed.

What to copy

Map the steps in your support chain and count what each one adds in delay and error. Move stable rules into the model and keep volatile facts outside it. Measure resolution without human escalation, and test the new design against the current one on real traffic.