Baidu has made a strategic pivot by open-sourcing its ERNIE 4.5 model family under the Apache 2.0 license. This marks a notable shift for the company, which previously kept its ERNIE models under wraps. It appears that rising competition from Chinese open-source contender Deepseek has spurred Baidu to open its doors to a global community of developers.
The company’s evolution in AI stretches back to its initial Ernie model, and now, with CEO Robin Li at the helm, it’s clear that Baidu wants to remove the barriers that often restrict building robust AI applications. Whether it’s dodging steep costs or avoiding proprietary lock-in, this move is designed to empower you as a developer.
The ERNIE 4.5 release includes ten different variants. Among these, you’ll find Mixture-of-Experts models boasting up to 47 active parameters and a total of 424 billion parameters, as well as a leaner 0.3 billion parameter option. To help you get started, Baidu has introduced developer tools like ERNIEKit and FastDeploy, making it easier to integrate these models into your projects.
Performance-wise, one standout model—the ERNIE-4.5-300B-A47B-Base—has outperformed Deepseek’s counterpart on 22 out of 28 benchmarks. While it hasn’t been directly compared with elite models like those from OpenAI, Deepseek’s R1, Claude 4, or Gemini 2.5 Pro, it certainly demonstrates strong capabilities in many areas.
Additionally, Baidu is testing what might be its answer to the R1 model in a new system dubbed X1, which, while not part of this launch, hints at further competition on the global AI stage. Industry voices, such as Sean Ren from the University of Southern California, have noted that when key labs open-source powerful models, it prompts the whole industry to aim higher.
In response to these shifts, OpenAI’s CEO Sam Altman has mentioned plans to roll out an open-source release in the future. If you’ve ever struggled with restrictive APIs or premium pricing models, this newer approach might just make a welcome difference.
