ERNIE 4.5 VL 28B A3B Thinking

ERNIE-4.5-VL-28B-A3B-Thinking upgrades Baidu’s ERNIE VLM with stronger visual reasoning, grounding, and STEM problem solving. It trains on large visual-language reasoning corpora, then stabilizes MoE with GSPO and IcePop strategies, plus dynamic difficulty sampling for efficient learning. The model can zoom into images, call tools like image search for long-tail facts, and handle multi-image, chart, and video understanding while keeping answers grounded. It offers 128k context, Apache 2.0 licensing, and production paths via transformers, vLLM, and FastDeploy, with fine-tuning recipes in ERNIEKit. Inference typically runs with about 3B active parameters per token, which helps balance quality and latency in enterprise multimodal assistants.