How ERNIE‑4.5‑VL Redefines Multimodal AI with 100+ Language Support

Introducing ERNIE‑4.5‑VL

When artificial intelligence enters a golden era of deep application, the limits of single‑modality are being shattered by multimodal interaction. The ERNIE‑4.5‑VL visual‑language model (ERNIE‑4.5‑VL‑28B‑A3B; ERNIE‑4.5‑VL‑424B‑A47B) delivers breakthrough image, video, and text understanding and reasoning, bridging the digital and physical worlds and supporting interaction in over 100 languages.

Performance Highlights

Experimental results show that the lightweight ERNIE‑4.5‑VL‑28B‑A3B dramatically reduces activation parameters, yet remains competitive—and often superior—in most benchmark tests when compared with models such as Qwen2.5‑VL‑7B and Qwen2.5‑VL‑32B.

Scalability and Modes

ERNIE‑4.5‑VL supports a 128K context length and offers two operating modes: a “thinking” mode for deep problem solving and a “non‑thinking” mode for rapid responses to basic tasks, making it flexible for everyday scenarios and professional domains.

Core Multimodal Capabilities

The model’s cross‑modal abilities cover a range of key task scenarios.

Related Resources

Wenxin Large Model Technical Blog (including reports): https://yiyan.baidu.com/blog/posts/ernie4.5

Hugging Face – Baidu models: https://huggingface.co/baidu

PaddlePaddle Star River Community: https://aistudio.baidu.com/modelsoverview?sortBy=weight&q=ernie

GitHub – ERNIE repository: https://github.com/PaddlePaddle/ERNIE

ModelScope – ERNIE‑4.5‑VL‑28B‑A3B: https://modelscope.cn/models/dengcao/ERNIE-4.5-VL-28B-A3B-Paddle

ERNIEKit Documentation: https://github.com/PaddlePaddle/ERNIE/blob/develop/docs/erniekit.md

FastDeploy: https://github.com/PaddlePaddle/FastDeploy/