Wenmeng Zhou

Currently working on Model-as-a-Service and Application Components within Alibaba Cloud’s Baillain/Model Studio platform , a comprehensive development platform for large language models (LLMs). My responsibilities include: Leading feature development and optimization for model serving infrastructure supporting the Qwen series , Wan series , embedding models , etc. Focused on improving service stability , latency efficiency , and leveraging platform capabilities to achieve optimal inference performance. Designed and maintained APIs for fine-tuning workflows, supporting algorithms such as CPT , SFT , DPO , and RL-based methods like GRPO . Built end-to-end pipelines for training, deployment, and evaluation, providing best practices across multiple application scenarios. Contributing to core components including Agent builder , Workflow engine , and the App Gallery/Template system , enabling users to rapidly build intelligent applications with minimal effort. Led the integration of Model Control Plane (MCP) into Baillain, enabling cloud and custom deployments of industry-standard MCP servers, and bridging with Alibaba Cloud's broader API ecosystem. Driving internal efforts to develop a powerful full-code SDK , empowering developers to quickly prototype and deploy complex applications using standard coding practices, complementing the platform’s low-code capabilities.

As the Technical Lead of ModelScope, I am at the helm of developing the model ecosystem, toolchain, and application ecology for the ModelScope community. My team and I are dedicated to innovating and advancing open-source projects, notably the large model training and fine-tuning framework Swift, the large model performance evaluation framework LLMuses, and the large model application middleware ModelScopeAgent. In addition to ModelScope, I also oversee the technological development of the [DashScope](https://dashscope.aliyun.com/). This involves architectural design and development of core services like Qwen and Wanx. My role extends to providing standardized APIs and SDKs for Chat and Assistant services, as well as image and text generation. Additionally, I support the model training and assessment functionalities of the dashscope Model Service Platform through API and SDK development, ensuring a robust and scalable infrastructure.

Currently I am leading a CV team to build highly efficient framework for training and inference, and also developing SOTA algorithms for different areas in computer vision. We have been supporting almost all business teams in Alibaba group, such as search, recomendation, ads, alipay, Gaode, AE and so on. We have also support many clients of Alibaba Cloud in both public cloud and private cloud。 Our work have been opensourced, welcome to use: https://github.com/alibaba/EasyCV Additionally， Currently I am interested in federated learning, trying to implement different deep learning algothrims in federeated setting and bring them to business

Work on building a training platform for computer vision called EasyVision, which aims to provide an easy way to pretrain or finetune different cv models, including implementing SOTA algorithms in image classification, object detection, sematic segmentation, OCR metric learning and so on. We also build robust and highly efficient inference system in both online and offline scenario.

Mainly focus on OCR algorithms and bring ocr to business by building the ocr service architecture and product on Alibaba Cloud

work in large-scale learning group @IDST. Mainly focus on OCR using deep learning, object detection and image classification.

During this period, I worked with Wang Liang on a project that reconstruct the real world and let people travel inside.