ClinFusion

ClinFusion is a medical multimodal large language model from Alibaba DAMO Academy. It uses a compositional, cascaded vision encoder with a Cascade Spatial-Aware Locality Fusion operator to handle both 2D medical images and native 3D volumes such as CT within one fused encoder. It is released in 8B and 32B versions.

It targets visual question answering, radiology report generation, instruction following and text-only medical tasks. The authors report that it outperforms open-source medical MLLMs such as Hulu-Med and Lingshu on 20 of 24 benchmarks, and beats proprietary models including GPT-5.2 and Gemini-3-Flash on 13 of 16 multimodal benchmarks. These are self-reported results. A blinded evaluation by board-certified radiologists is also reported. The model can be extended with agentic tool use for retrieval-augmented clinical workflows.

The project also provides an evaluation suite: MedIF-Bench for instruction following, a region-of-interest-grounded report evaluation method, and the ClinFusion-Eval-Data dataset (211K records over 22 benchmarks, with a 509 GiB media archive). Code is on GitHub under Apache-2.0, weights are on Hugging Face, and an 8B demo runs in a Hugging Face Space. Installation requires a CUDA setup with a pre-built flash-attention wheel and several separately downloaded vision encoders. It is a research release and is not presented as a validated clinical product.