Show HN: 4B VLM trained on 600k questions to query images in one pass (github.com)

Dragonfly has emerged as a groundbreaking visual language model (VLM), fine-tuned from Qwen3-VL, designed to answer multiple questions about an image in a single pass while providing probabilistic outputs. Unlike traditional VLMs that generate responses sequentially for each question, Dragonfly processes the image only once, scoring each question independently. This innovation allows it to present answers in a more structured format with associated probabilities for each potential answer, enhancing both accuracy and interpretability. This development is significant for the AI/ML community as it offers a more efficient approach to visual question answering. By providing probability scores, Dragonfly not only improves performance metrics—such as accuracy in classification tasks—but also facilitates better understanding of model confidence. The model has been trained on a diverse dataset comprising over 600,000 human-labeled questions derived from popular visual question answering datasets. This positions Dragonfly as a strong competitor in the field, potentially leading to more robust and explainable AI applications in image analysis. Its implementation requires a single GPU with around 12 GB of memory, making it accessible for a range of users in the AI research community.