How Baidu’s New AI OS “WanYuan” Redefines Intelligent Computing

On April 16, Baidu held the Create 2024 AI Developer Conference in Shenzhen, where Baidu’s Executive Vice President and President of Baidu Intelligent Cloud, Shen Dou, announced the launch of a new intelligent computing operating system called Baidu Intelligent Cloud WanYuan. The OS is designed to manage clusters with tens of thousands of GPUs, fully exploit CPU/GPU performance, and embed powerful large‑model engines as core services.

Why a New OS Is Needed

Traditional cloud‑computing systems remain important but are no longer the main focus. The emerging “intelligent computing” platform requires a fresh abstraction layer that redefines human‑machine interaction and offers developers a simpler, smoother experience. WanYuan is positioned as that next‑generation OS.

Historical Context and Evolution

The speech traced OS evolution from early manual cable‑plugging, through assembly language, high‑level languages, and the rise of clusters. As software grew in size (UNIX 6 at 9 k lines, Windows NT at 4.3 M lines, Windows XP at 45 M lines), operating systems evolved from managing single‑process machines to orchestrating micro‑services across clusters.

Core Architecture

WanYuan’s architecture consists of three layers:

Kernel Layer : Manages heterogeneous hardware (GPU‑centric), integrates large models (ERNIE 4.0, 3.5, Speed/Lite/Tiny, Wenxin Vision, third‑party models), and provides a unified interface for AI‑native applications.

ModelBuilder Layer : Handles model management, scheduling, fine‑tuning, and routing. It enables developers to adapt existing models with minimal data, achieving up to 30% inference cost reduction while maintaining performance.

Tool Layer : Includes Qianfan AppBuilder and AgentBuilder for rapid application development, workflow orchestration, and one‑click deployment to various platforms.

Performance Highlights

On Baidu’s internal “Bai Ge” platform, WanYuan achieves 98.8% effective training time utilization on a ten‑thousand‑card cluster, with linear acceleration and bandwidth efficiency exceeding 95%. Mixed‑chip training across different vendors incurs only a 3% performance loss at ten‑thousand‑card scale (≤5% loss at thousand‑card scale). This demonstrates the OS’s ability to hide hardware heterogeneity and maximize cluster efficiency.

Ecosystem and Future Plans

The OS is built to be open and extensible. Baidu plans to invite partners to contribute models, tools, and applications, further enriching the WanYuan ecosystem. Future work includes expanding chip support, simplifying heterogeneous cluster management, and providing stable, secure, high‑performance intelligent computing platforms for vertical industries.

Overall, WanYuan aims to make AI development as simple as building with blocks, turning every idea into a deployable application within minutes.