MedChain-R1-9B is a 9B vision-language model for sequential medical image grounding, built on Qwen3.5-9B and described in "Think Before You Locate: Clinical Chain-of-Thought Supervision for Sequential Medical Image Grounding" (arXiv:2611.00041). Given a sequence of medical images and a clinical question, it first writes an image-grounded rationale in <think> tags, then returns one pixel-coordinate bounding box. Training used deliberative chain-of-thought supervision (dCoT-SFT) on the MedChain corpus, followed by GRPO with an IoU plus format reward. MedChain has 182,573 filtered rationales over the 8 MedSG-Bench tasks and 10 imaging modalities. A medical VLM (FlemingVL-38B) drafts each rationale and a multimodal refiner (Qwen3.5-397B-A17B) rewrites it conditioned on the annotated box. This checkpoint scores 77.54 mIoU / 84.87 Acc@0.5 on all 9,630 MedSG-Bench test questions, above the paper's 9B GRPO row (76.90 / 83.82), which came from a different checkpoint. It also beats the supervised state-of-the-art MedSeq-Grounder 7B (72.55 / 79.71) and is far ahead of the zero-shot general and medical baselines, which mostly score below 21 mIoU. The paper notes that the gains vary by task. Against an answer-only 4B model, the reasoning supervision gains 14.7–16.1 mIoU on difference grounding but loses 2.2–3.2 on multi-view, patch and referring grounding. It adds that GRPO contributes little after dCoT-SFT (+0.37 mIoU at 4B). The card says to generate one example at a time, because batched generation gave degenerate outputs, and to use the bundled processor. The rationales were written knowing the answer, so they may be post-hoc rather than a faithful record of the model's reasoning. The model is a research artifact, not validated for diagnosis or clinical deployment, and evaluation is limited to MedSG-Bench. It is released under Apache 2.0.
docker model run hf.co/prithivMLmods/MedChain-R1-9B-GGUF: