Poly Hotels Details Vertical LLM Path, Lightweight Model Deployment Under RMB 100,000

Key points

• Zhang Ye, deputy general manager of Poly Hotels Management, said AI is widespread in shallow hotel scenarios such as front-desk service and daily operations, but limited in core decisions like revenue growth, operational coordination and strategy.
• Zhang argues the hotel industry's low-price, high-frequency traits make general LLMs a poor fit, and uploading private data such as pricing records, footfall and revenue to public clouds risks leaks.
• Poly's approach applies hotel-specific reinforcement learning on a general model base, injects an industry knowledge base, then distills and compresses it into a lightweight small model.
• The lightweight model retains over 95%-98% of the original vertical model's core capabilities, with deployment cost of only RMB 200,000, or even under RMB 100,000.
• Poly says its model, training software, data-cleaning software and chips are all domestically produced, with a Harness framework wrapping the model to run agents and Skills.
• The model is in internal testing for business analysis, monitoring over 1,000 metrics and issuing alerts on more than 170 first-tier indicators, with external release planned for October or November.
• Poly says it will package data-governance, training and performance-evaluation tools for small and mid-sized hotel managers, with a model of self-purchased hardware plus monthly service fees in the thousands of RMB.