Chinese tech giant Baidu unveiled new AI products at its Wave Summit conference held on September 9, including the ERNIE X1.1 reasoning model, an upgraded PaddlePaddle framework and a new version of its coding assistant Comate. The company also open-sourced a new ERNIE-4.5 model designed for advanced reasoning tasks.
Baidu said ERNIE X1.1 delivers improvements in factuality, instruction following and agentic capabilities. Compared to its predecessor, ERNIE X1, the new version showed a 34.8% increase in factual accuracy, a 12.5% gain in instruction following and a 9.6% rise in agentic functions.
The company stated that benchmark results showed ERNIE X1.1 outperforming DeepSeek-R1-0528 in overall performance and matching leading global systems, such as GPT-5 and Gemini 2.5 Pro.
Built on Baidu’s multimodal ERNIE 4.5 foundation model, ERNIE X1.1 uses a hybrid reinforcement learning framework combining mixed reinforcement learning and iterative self-distillation.
The model is now available through the ERNIE Bot website, Wen Xiaoyan app and Baidu AI Cloud’s Qianfan platform for enterprise clients and developers.
At the event, Baidu also released PaddlePaddle framework version 3.2, which features computational optimisation, parallel strategies and native fault tolerance. According to the company, the framework achieved a 47% machine FLOPs utilisation rate on large-scale pre-training. The PaddlePaddle-ERNIE ecosystem has so far served 23.33 million developers and 7,60,000 enterprises.
The firm further introduced Baidu Comate 3.5S, an upgraded AI coding assistant with expanded multi-agent collaboration. The system now allows multiple AI agents to work together on complex tasks and supports integration of unified knowledge bases for team collaboration.
“Within Baidu, 45% of new code is already generated by AI,” the company said, adding that the tool has supported over 10 million developers.
In addition, Baidu open-sourced ERNIE-4.5-21B-A3B-Thinking, a reasoning model with 21 billion total parameters and 3 billion active parameters. The model supports a 128K context window and is designed for applications such as content generation, reasoning, mathematics, code and tool use.
The company said it offers near state-of-the-art performance at reduced scale and is now available on Hugging Face and Baidu AI Studio.
