Medical AI is constrained by fragmented, privacy-sensitive datasets and the high cost of expert annotation, making brute-force scaling impractical. This project proposes MMBS (Medical Model Beyond Scaling), a strategy-driven framework that builds robust medical imaging foundation models without relying on billion-scale datasets. Our core idea is “strategy over scale”: we systematically combine algorithmic innovations across representation learning, fine-tuning, and modular inference to maximize transfer under limited medical data. At the representation level, we fuse two complementary self-supervised learners, MAE (capturing global anatomical structure via reconstruction) and DINO (capturing local semantic alignment), producing a unified embedding suitable for diverse clinical tasks. At the adaptation level, we implement multi-stage fine-tuning (MSFT), intermediate supervised fine-tuning (ISFT), and parameter-efficient fine-tuning (PEFT) to strengthen domain knowledge while preserving generalization and robustness across institutions. At the inference level, we design a Transformer-style decoder that supports a unified multi-task pipeline spanning detection, segmentation, and classification, enabling interpretable outputs and alignment with clinical reasoning. Validation follows a progressive cross-domain strategy: we start with fundus imaging and expand to visually challenging modalities such as gastroscopy and echo, in close collaboration with clinical teams at Chang Gung Memorial Hospital. Deliverables include reproducible training/evaluation assets and a cross-domain medical imaging AI toolkit to support model sharing, clinical translation, and trustworthy deployment within Taiwan’s medical AI ecosystem.