Show HN: 4B VLM trained on 600k questions to query images in one pass

Heyo!

I'm Adit, an undergrad in Finland. This is a preview of a side project I worked on last weekend. Essentially, Dragonfly is a LoRA fine-tune of Qwen3-VL with typed output heads that answers many questions about an image in one pass, with a probability for every answer. Unlike a VLM that generates an answer to each question, Dragonfly reads the image once and scores every question in the same pass. Questions can't see each other, so each answer is the same as asking it alone, in any order.

Instead of text outputs, Dragonfly answers with calibrated probabilities for the options you give it in order to answer the question. You can ask Dragonfly multiple choice, yes/no, or scale-based questions.

As you might have guessed already, I got the idea from Jev and thought it would be fun to train a small model that can do the same thing for images. There wasn't a dataset for questions specifically, so I compiled a couple open datasets to form around half of our training data and then distilled Qwen to generate ~300k questions that I could verify against the human annotated data.

Try it out @ https://adit.run/dragonfly and lmk what you think :)