Customer Service Automation
Quwan Technology needed a cheaper and faster way to train, tune, and deliver realistic digital humans because traditional workflows took days, suffered from poor lip sync and expression quality, and made enterprise deployment hard to scale.
Reported impact
Quwan reduced digital-human delivery time and cost, improved visual and speech quality, and expanded enterprise deployments across telecom and other industries.
Quwan Technology built a digital-human production line on Huawei Cloud so its self-developed multimodal models could be turned into enterprise service agents and media assets. Ascend AI Cloud Service provides the compute base for large-model training and inference, while ModelArts Studio manages tuning, compression, deployment, evaluation, elastic resource allocation, checkpoint recovery, and iteration. Quwan uses this stack to refine face, speech, and expression models so digital humans can better hear, speak, think, and respond in business scenarios. The same pipeline supports rapid delivery of photorealistic, multilingual avatars and powers deployments such as China Telecom's 'Xiaoxin' digital assistant, which combines real-time recognition, dialogue, perception, and generation for one-stop service handling. By linking training compute, model tooling, and business delivery, Quwan shortened model delivery from months to weeks.
Model training and inference run on Ascend cloud resources, while ModelArts Studio manages tuning, deployment, and recovery. The trained digital-human models are then served through Quwan's production line into industry-specific assistants such as Xiaoxin, where real-time recognition, dialogue, perception, and generation are combined for user interactions.
"A single digital human traditionally took 3-7 days and still might have lip-sync and expression problems. (original: 一般一个数字人制作时间长达3-7天，制作完成后可能存在口型对不上，表情不精准等问题。)"
"Ascend Cloud Service kept model training uninterrupted for 30 days, and lip-sync accuracy exceeded 95% while the expression MOS score exceeded 4.0 out of 5. (original: 基于昇腾云服务的长稳能力，模型训练30天不中断，加速了大模型的快速迭代；华为云提供昇腾算力助力趣丸进行人脸模型算法优化，数字人口型匹配率大于95%，远超业界水平，生成的数字人表情质量主观MOS分超过4.0（满分5.0）。)"
"ModelArts Studio raised overall development efficiency by more than 50% and reduced large-model delivery time from months to weeks. (original: ModelArts Studio的联合协同调优能力，让趣丸与同高效师生可进行联合调优，提升整体开发效率50%以上；... 大模型交付时长从月到周。)"
"Quwan built a digital-human production line on Huawei Cloud Ascend AI Cloud Service and ModelArts Studio, making it possible to generate ultra-realistic digital humans in minutes. (original: 趣丸科技和华为云合作，通过华为云昇腾AI云服务为大模型等提供充沛算力支持，加速模型的训练和推理，通过ModelArts Studio提升算法调优能力，简化研发流程，缩短模型迭代周期，实现分钟级即可生成超写实数字人。)"
"Quwan's Xiaoxin digital human for China Telecom combines real-time recognition, dialogue, perception, and generation to guide users through one-stop service consultation and handling. (original: 趣丸科技为中国电信营业厅打造的数字人“小信”，融合了实时智能识别、智能对话、智能感知、智能生成等多项人工智能核心技术，让数字人富有拟人化、情感化，可通过实时语音对话，智能引导用户完成“一站式”的业务咨询及服务办理。)"
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Huawei Cloud Sample Archived · Dec 14, 2024 Kuakua Jingling (夸夸菁领) Kuakua Jingling needed a secure, always-on platform for AI digital employees that could handle sensitive government and enterprise workflows while supporting fast model iteration and tenant-specific knowledge. 7x24 automated execution capability DDoS protection service · ModelArts · Pangu large model 
Google Cloud Sample Modified · Oct 9, 2026 Tally Group This rapid processing surfaces targeted insights in under 20 seconds, eliminating manual checklists. At utility billing platform Tally Group , call center agents previously navigated up to six applications to… Apigee API Management · BigQuery · Gemini platform 
Anthropic Sample Archived · Oct 8, 2026 Zendesk Zendesk wanted customer-specific service agents to handle the complex work between a request and its resolution, but one-off automations were expensive to build, hard to localize, and difficult to govern across millions… 1 million custom agent executions in the first seven weeks of early access Amazon Bedrock · Claude Agent SDK · Claude Code 