.svg){kind=spacer}
Details
- Baidu has introduced ERNIE 5.0, the newest member of its Wenxin large model series, officially unveiled on November 13, 2025.
- The model features a 2.4-trillion-parameter Mixture-of-Experts architecture but activates less than 3 percent of its experts per inference, significantly reducing GPU demands and latency.
- Native omni-modal training enables a single model to understand and generate text, images, audio, and video without external adapters.
- Baidu claims this unified approach streamlines prompt engineering and allows for fluid multi-modal reasoning, such as describing images, responding to questions, and generating audio—all in one interaction.
- ERNIE 5.0 will support products like Ernie Bot, Baidu Search, and intelligent cloud APIs, with an early access SDK available to enterprise customers this quarter.
- Billed as the successor to ERNIE 4.0, the new model reportedly achieves double the benchmark efficiency per token on Nvidia H100 clusters.
- Baidu has provided a technical paper and demo site for researchers and developers to explore ERNIE 5.0's capabilities.
Impact
This launch positions Baidu closer to global leaders like OpenAI and Google as the competition heats up around massive sparse expert models. The model’s efficiency and broad modality coverage may significantly lower costs, making leading-edge AI more accessible to Chinese enterprises, while aligning with China’s push for sustainable AI development. Rival offerings from Alibaba and Huawei could accelerate as a result, potentially reshaping China’s cloud AI landscape in both research and deployment.
