ERNIE 5.0 Thinking Preview is a natively multimodal foundation model developed by Baidu, unveiled in November 2025 at the Baidu World event. As the latest generation of the Wenxin series, the model is designed with a "system-2" reasoning approach that incorporates extended processing time—often referred to as "thinking-time"—to handle complex logic, multi-step planning, and agentic tasks. It represents a transition toward unified modeling, where different data types are processed within a single coherent framework.
The model utilizes an ultra-sparse Mixture-of-Experts (MoE) architecture with approximately 2.4 trillion total parameters. To optimize inference and reduce computational costs, it employs modality-agnostic expert routing that activates less than 3% of its parameters for any given task. This architecture allows the model to jointly model text, images, audio, and video from the ground up, moving away from the late-fusion methods used in earlier multimodal systems.
Key capabilities of the Thinking Preview variant include advanced factual reasoning, creative writing, and autonomous tool use. It is specifically optimized for high-cognitive-load scenarios such as scientific deduction, mathematical proofing, and complex software engineering. The model is accessible through Baidu's Wenxin ecosystem for both individual consumers and enterprise developers.