- ERNIE X1.1 shows major advancements in factuality, instruction following, and agentic capabilities; it surpasses DeepSeek R1-0528 in overall performance while performing on par with top-tier models like GPT-5 and Gemini 2.5 Pro.
- The PaddlePaddle-ERNIE ecosystem now serves 23.33 million developers and 760,000 enterprises; while the AI coding assistant Baidu Comate has supported more than 10 million developers.
- Baidu also announced the open-source of ERNIE-4.5-21B-A3B-Thinking model.
BEIJING, Sept. 9, 2025 /PRNewswire/ — At WAVE SUMMIT 2025, the company’s deep learning developer conference, Baidu unveiled a series of AI advancements that further accelerate the wide adoption of AI. These advancements include the company’s latest reasoning model ERNIE X 1.1, an upgrade to the PaddlePaddle framework, the upgraded AI coding assistant Baidu Comate 3.5S, as well as the open-sourcing of the ERNIE-4.5-21B-A3B-Thinking model.
ERNIE X1.1 Shows Major Advancement in Factuality, Instruction Following, and Agentic Capabilities
The newly released ERNIE X1.1 reasoning model is a significant upgrade that delivers major advancements across core capabilities including factuality, instruction following, and agentic capabilities. Compared to ERNIE X1, ERNIE X1.1’s factuality is up 34.8%, instruction following is up 12.5%, and agentic capabilities are up 9.6%
Multiple benchmark results show that ERNIE X1.1’s overall performance surpassed that of DeepSeek R1-0528, demonstrating clear advantages in certain tasks. Meanwhile, its performance is on par with other top-tier models such as GPT-5 and Gemini 2.5 Pro. Built upon ERNIE 4.5, a multimodal foundation model, ERNIE X1.1 has incorporated an iterative hybrid reinforcement learning framework that blends mixed reinforcement learning and iterative self-distillation.
ERNIE X1.1 is now accessible to users via the ERNIE Bot website (ernie.baidu.com) and the Wenxiaoyan app, and has also been deployed on Qianfan, Baidu AI Cloud’s MaaS (Models-as-a-Service) platform — open to enterprise clients and developers.
PaddlePaddle Framework v3.2 and Baidu Comate 3.5S Lower Barriers for AI Development and Deployment
At the event, Baidu also debuted the PaddlePaddle framework version 3.2, featuring enhancements in model training, hardware compatibility, and ecosystem support. The upgraded PaddlePaddle framework incorporates three core upgrades—computational optimization, parallel strategies, and native fault tolerance—that address long-standing challenges in model training. The framework significantly boosts efficiency, achieving a pre-training MFU of 47% on the ERNIE-4.5-300B-A47B model.
Accompanying the new framework are a suite of toolkits, including ERNIEKit, designed for foundation model development, FastDeploy v2.2 for efficient model deployment, and two scientific computing toolkits, PaddleCFD and PaddleMaterials. Together, these releases further reduce barriers to applying foundation models across training and deployment.
To date, the PaddlePaddle-ERNIE ecosystem has served 23.33 million developers and 760,000 enterprises.
Also released at the event was Baidu Comate 3.5S (Super Synergistic AgentS), a major upgrade to its AI coding assistant. The new version strengthens multi-agent-collaboration capabilities, enabling a single developer to achieve the productivity and efficiency of an entire team.
Baidu Comate 3.5S delivers three core enhancements. First, agents are more powerful, with stronger business understanding, the ability to break down task lists, and improved execution accuracy. Second, the system advances from single-agent to multi-agent collaboration, dynamically generating multiple agents that autonomously work together to complete complex tasks. Third, team collaboration is reinforced through the integration of Rules and MCP, creating a unified knowledge base that enables seamless knowledge transfer within teams.
To date, Baidu Comate has already supported more than 10 million developers. Within Baidu, 45% of new code is now generated by AI.
Open-sourced ERNIE-4.5-21B-A3B-Thinking Model Delivers Outstanding Performance Across Tasks
At the conference, Baidu announced the open-sourcing of ERNIE-4.5-21B-A3B-Thinking, a deep-thinking model trained on the foundation of ERNIE-4.5-21B-A3B. It supports a 128K context window and is designed for complex reasoning tasks that require long-context understanding. The model delivers outstanding performance across a wide range of tasks, including content creation, logical reasoning, mathematical computation, code generation, and tool use.
As a lightweight MoE model with 21 billion total parameters and only 3 billion active parameters, ERNIE-4.5-21B-A3B-Thinking has achieved performance close to leading large-scale models across benchmarks, delivering near-SOTA intelligence at a significantly reduced scale.
ERNIE-4.5-21B-A3B-Thinking is now available on platforms including Hugging Face and Baidu AI Studio. The release of this model follows the open-sourcing of 10 other distinct ERNIE 4.5 models on June 30.
About Baidu
Founded in 2000, Baidu’s mission is to make the complicated world simpler through technology. Baidu is a leading AI company with strong internet foundation, trading on the NASDAQ under “BIDU” and HKEX under “9888.” One Baidu ADS represents eight Class A ordinary shares.
Media Contact
[email protected]
