Alibaba's ClinFusion: 8B Medical VLM Beats Gemini 3 Flash on Most Benchmarks
jiqizhixin · x · 2026-08-29
Medical VLMs face a dilemma: 2D-centric models lose 3D structure while 3D-centric ones need expensive alignment, and existing benchmarks don't measure what clinicians care about. Alibaba DAMO Academy's ClinFusion (8B and 32B) tackles both.
Its core is CaSL Fusion, a local cross-attention that fuses 2D and 3D features without increasing visual tokens sent to the LLM — raising per-token information density while cutting compute from quadratic to linear. It also introduces a clinically aligned, factuality-oriented evaluation framework.
Results: first or second on all eight 2D VQA benchmarks, with the 8B beating Gemini-3-Flash on most; on abdominal CT QA it outperforms the strongest open-source same-size model by 14.5 points.
More from Multimodal
- SEEDANCE generates realistic DC travel vlog: Detailed prompt breakdown — SimplyAnnisa · 2026-08-29
- Minimax H3 Max generates video faster than real-time, powering infinite 'Interdimensional Cable' Twitch stream — menhguin · 2026-08-29
- Midjourney sref code yields warm, bright painterly sunny vibes — op7418 · 2026-08-29
- Fun test: Asking AI to generate a deeply wrong reality — aerodynamic_werewolf · 2026-08-29
- MiniMax H3优化实测：Turbo 模式提速 2 倍 — BluePointDigital · 2026-08-29
- Entire clip generated in 2 minutes with H3 Max and Nano Banana Pro via fal agent — isidentical · 2026-08-29