Baidu Unveils ERNIE X1.1 with Enhanced Capabilities

Baidu has introduced its latest reasoning model, ERNIE X1.1, showcasing significant improvements in factuality, instruction following, and agentic capabilities, as announced in a press release. The model outperforms DeepSeek R1-0528 and matches top-tier models like GPT-5 and Gemini 2.5 Pro in overall performance.

ERNIE X1.1, built upon the ERNIE 4.5 multimodal foundation model, incorporates an iterative hybrid reinforcement learning framework. This framework blends mixed reinforcement learning with iterative self-distillation, enhancing the model's capabilities. The model is now accessible via the ERNIE Bot website and the Wenxiaoyan app, and is also available on Qianfan, Baidu AI Cloud's Models-as-a-Service platform.

In addition to ERNIE X1.1, Baidu unveiled the upgraded PaddlePaddle framework version 3.2 and the AI coding assistant Baidu Comate 3.5S. The PaddlePaddle framework features enhancements in model training and hardware compatibility, while Baidu Comate 3.5S strengthens multi-agent collaboration capabilities. Baidu also announced the open-sourcing of the ERNIE-4.5-21B-A3B-Thinking model, designed for complex reasoning tasks, now available on platforms like Hugging Face and Baidu AI Studio.

Baidu, Inc. continues to expand its AI ecosystem, serving millions of developers and enterprises globally.