Baidu Open-Sources ERNIE Bot 4.5 Series, Accelerating AI Development

Beijing, June 30, 2025 — Baidu has officially open-sourced its ERNIE Bot 4.5 series, marking a significant milestone in China's large model ecosystem. The release includes ten models, featuring a 47B/3B mixture-of-experts (MoE) model and a 0.3B dense model, alongside full pre-training weights and inference code.

Availability and Deployment

The models are now accessible on:

• PaddlePaddle Starry Community
• Hugging Face
• Baidu Intelligent Cloud's Qianfan Platform (via API services)

This strategic move positions Baidu alongside Tencent, Alibaba, and ByteDance as key contributors to China's open-source AI landscape.

Image source note: The image is AI-generated, and the image licensing service provider is Midjourney.

Technical Specifications and Performance

The ERNIE Bot 4.5 series demonstrates:

• Native multi-modal capabilities (text, images, videos)
• Superior benchmark performance vs. GPT-4o in select tests
• Optimized for resource-constrained environments (3B/0.3B models)

Developers highlight the smaller models' efficiency for edge devices, potentially rivaling DeepSeek V3 and Alibaba Qwen in specific use cases.

Core Innovations

1. Multi-modal MoE Architecture: Simultaneously processes text and visual data streams
2. High-Efficiency Infrastructure: Enables rapid training/inference cycles
3. Modality-Specialized Post-Training: Enhances task-specific performance

Market Impact

The release intensifies competition in the global AI market, pressuring closed-source providers while:

• Raising industry technical standards
• Empowering developer communities
• Accelerating AI application development cycles

Key Points

• Baidu releases 10 open-source models including MoE and dense variants
• Full-weight availability on major AI platforms
• Demonstrates multi-modal superiority over competitors like GPT-4o
• Targets both cloud (47B) and edge (0.3B) deployments
• Advances China's position in the global AI race