ERNIE-4.5-VL 28B-A3B Thinking
ERNIE-4.5-VL 28B-A3B Thinking is an Apache-2.0 multimodal model from Baidu that’s positioned around visual-language reasoning: chart and diagram understanding, STEM-style problems from screenshots/photos, and more structured visual grounding. The model card frames it as an MoE architecture where only a smaller subset of parameters is activated at inference time, and it highlights mid-training plus reinforcement-learning style tuning on verifiable tasks.
What to try first: run the provided Transformers quickstart on a small “describe/ground this image” prompt, then try a tougher input like a multi-step chart question or a geometry problem from a photo. If your workflow depends on tool use (for example, image search or zoom/crop flows), the model card explicitly calls out those behaviors—so it’s worth testing whether they work end-to-end in your own wrapper before committing to a larger integration.
Quick stats from the listing feed: pipeline: image-text-to-text · 518 likes · 1048 downloads.
Source listing: https://huggingface.co/models?sort=modified