58.com’s LingXi Large Language Model Platform: Development, Deployment, and Performance Optimizations

Following the release of ChatGPT on November 30, 2022, 58.com adopted a Model‑as‑a‑Service (MaaS) approach and launched a large‑language‑model (LLM) platform in May 2023 that provides training, inference, and one‑stop development services for internal developers.

Building on open‑source general LLMs and leveraging data from the company’s life‑service domain, 58.com continued post‑training to create a vertical LLM named LingXi ChatLing, and in 2023 began applying the technology to several AI products.

In 2024 the goal is to build a leading, agile, and easy‑to‑use AI platform that accelerates AI application deployment across the company, achieved by enhancing platform capabilities, deepening collaboration with business units, upgrading existing services with LLMs, and creating benchmark applications.

By the end of 2024 the platform supports more than 100 internal business scenarios, handles over ten million daily inference calls—a near‑twenty‑fold increase from the start of the year—and the LLM technology has permeated sales, customer service, product, operations, and internal workflows.

Key flagship applications built on the LingXi model include AI interview assistants, AI housing search, AI chat agents, AI content review, intelligent outbound calling, intelligent customer service, sales assistants, smart office assistants, big‑data assistants, and AI code assistants, as well as early deployments in recruitment search and real‑estate recommendation.

To sustain the growing traffic, the team has focused on inference performance and GPU utilization by integrating vLLM/SGLang/LMDeploy frameworks, applying quantization, upgrading model versions (doubling performance), using S‑LoRA and vGPU for mixed‑GPU deployment, and developing lightweight ChatLing‑Mini models through pruning and knowledge distillation; they also introduced ModernBERT, fine‑tuned on domain data, which outperforms generative LLMs in many NLP tasks.

To lower the barrier for AI application development, the LingXi Intelligent Agent platform was built on top of the LingXi model, offering RAG, tool calling, and workflow capabilities with zero‑code configuration and auto‑generated APIs; role‑play and function‑call abilities were enhanced, and a domain‑specific text‑embedding model (WTE) was trained for hybrid retrieval.

The platform’s LLM was evaluated on the LiveBench benchmark, achieving a score of 57.52 and ranking among top‑tier models, with the API already submitted to LiveBench for official inclusion.

In parallel, the team incorporated open‑source voice LLMs and extensive voice data to improve voice generation quality, applying the technology to dialogue systems, video generation, and digital humans, and combined knowledge distillation with streaming generation to enable low‑latency, end‑to‑end conversational solutions for intelligent outbound calls and digital‑human interview scenarios.

Overall, by 2024 more than a hundred business scenarios within 58.com have integrated the large‑language‑model platform, with daily inference exceeding ten million calls, demonstrating broad AI adoption while acknowledging the long investment horizon and uncertainty inherent in AI‑driven business transformation.

2025 New Year greetings from Zhan Kunlin, Senior Director of 58.com AI Lab.