DownloadERNIE-4.5-VL-424B-A47B-Paddle is a powerful multimodal model designed by Baidu, integrating both vision and language capabilities through a MoE (Mixture of Experts) architecture. With 424 billion total parameters and 47 billion activated per token, it is optimized for complex image-text understanding, reasoning, and generation. Post-trained using supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and alignment strategies, the model is capable of toggling between thinking and non-thinking modes during inference. Built on PaddlePaddle, it supports fast deployment with FastDeploy and scalable parallel inference. The model's heterogeneous MoE structure ensures distinct learning for both modalities without interference, reinforced by modality-specific routing and balanced token loss. It excels in multimodal chain-of-thought reasoning, image captioning, and instruction-following tasks.
Features
- 424B parameters with 47B activated per token
- Supports text and image modalities simultaneously
- RLVR and SFT used for alignment and reasoning enhancement
- Built-in thinking mode toggle for flexible inference behavior
- PaddlePaddle-based, with FastDeploy and vLLM compatibility
- Handles long contexts up to 131,072 tokens
- Uses modality-isolated routing and token-balanced loss
- Supports 4-bit and 8-bit quantized deployment for efficiency
Project Samples
