CUHK-Tsinghua MedJev Strips Autoregressive Tail Off Qwen3.5-4B, Posts 7.93-Point Medical VQA Gain at 6× Speed
Summary
MedJev, from CUHK, Tsinghua and NUS, is a 'system one' vision-language decision model that scores candidate answers via pretrained query-key matching rather than generating tokens. Built on Qwen3.5-4B with its top language layers removed, MedJev-4B posts 7.93 absolute points of accuracy gain and 6.04× end-to-end speedup over the base model across four medical image classification and VQA benchmarks, extending the fast-decision-model trend into radiology.
Originally reported by huggingface.co
Read the original article →Original headline: CUHK-Tsinghua MedJev Strips Autoregressive Tail Off Qwen3.5-4B, Posts 7.93-Point Medical VQA Gain at 6× Speed