How Baidu’s Wenxin Model Hit 430 Million Users and What Its New Tech Means for AI

During Baidu World 2024 in Shanghai, Baidu’s chief technology officer Wang Haifeng announced that the Wenxin Yiyan large‑model service has amassed 430 million users and that daily calls to the model exceed 1.5 billion, underscoring its massive adoption.

Continuous Model Evolution

Since its launch on March 16, 2023, Baidu’s Wenxin series has undergone several upgrades, adding knowledge‑enhancement, retrieval‑enhancement, and intelligent‑agent capabilities. The retrieval‑augmented approach follows a "understand‑retrieve‑generate" pipeline that jointly optimizes these stages, improving both the quality and relevance of generated content.

Retrieval‑Augmented Multimodal Generation

For text‑to‑image tasks, Baidu introduced a retrieval‑augmented generation technique that can produce both highly precise images (exact visual replication) and generalized artistic renditions. The workflow first analyzes user intent, decides whether a precise or generalized output is needed, retrieves relevant reference images, and then generates the final picture using a multimodal controllable model. Local‑attention computation preserves entity features while enabling high‑level style variations (e.g., turning Newton’s portrait into a story‑book illustration), whereas global‑attention computation ensures pixel‑perfect fidelity for exact matches such as reproducing a specific car.

Intelligent Agents and Multi‑Agent Collaboration

In October 2023, Baidu unveiled its intelligent‑agent framework, dubbed System 2, which incorporates a "thinking model" comprising understanding, planning, reflection, and evolution stages. This framework enables reliable execution, self‑evolution, and partial white‑box transparency, allowing machines to reason and act like humans on complex tasks.

Complex tasks often require collaboration among multiple agents. Baidu highlighted three collaboration patterns—centralized, decentralized, and pipeline—to boost overall system efficiency and problem‑solving capability.

Agent‑Powered Coding Tools

The "Miaoda" (秒哒) tool adopts a centralized‑agent architecture, where a code‑generation agent enriched with scenario knowledge produces higher‑quality, more stable code. The "Wenxin Fast Code" (文心快码) platform, now at version 3.0, demonstrates a full development pipeline: agents handle requirement analysis, code synthesis, testing, and bug fixing, dramatically improving developer productivity.

Model Matrix and Platform Integration

Baidu’s model portfolio includes flagship models such as ERNIE 4.0 Turbo and lightweight variants like ERNIE Speed, alongside specialized thinking and scenario models built on top of the base models. These are tightly integrated with the PaddlePaddle deep‑learning platform, delivering gains in model performance, training efficiency, and inference speed.

Key statistics disclosed at the event: PaddlePaddle now has 18.08 million developers, serves 430 thousand enterprises, and hosts 1.01 million trained models.

Industry Applications

AI is increasingly permeating various sectors. Baidu partnered with the China Science and Technology Museum to deploy AI‑driven exhibits and a "museum intelligent agent" that enhances public science education. In sports, an AI‑assisted diving training system combines cloud, edge, and device technologies to automatically capture, track, and analyze each dive, providing real‑time scoring and detailed performance metrics based on the Wenxin model.

These initiatives reflect Baidu’s broader strategy of applying large‑model capabilities to real‑world problems across education, entertainment, and professional training.

Talent Development and Future Outlook

To support the AI ecosystem, Baidu reported that it has cultivated 5.92 million AI‑related talents for society. Wang Haifeng concluded by affirming that the Wenxin model continues to be trained, with more powerful versions on the horizon.