Pocket-Dentist

Dental image understanding is emerging as an important application area for VLMs, but current evaluation remains fragmented across modality-specific datasets, isolated task definitions, and metrics that rarely account for deployment constraints. As a result, it remains unclear whether VLMs can provide reliable dental image understanding across heterogeneous modalities and tasks while remaining practical for resource-constrained deployment. We present Pocket-Dentist, a large-scale multimodal benchmark and deployment pipeline for dental VLMs. We curate and standardize seven dental datasets into a unified vision–language benchmark comprising more than 6,000 patients, 71,000 images, four imaging modalities, six task types, and 14 evaluation metrics. On this benchmark, we evaluate 14 VLMs under zero-shot, few-shot, and LoRA fine-tuning settings, including 12 open-weight models under a uniform LoRA adaptation budget. The results show that zero-shot and few-shot performance is fragmented across modalities and tasks, whereas dental-domain LoRA adaptation enables compact VLMs to become competitive with substantially larger models. In particular, Qwen3-VL-4B achieves the strongest overall performance among compact models and matches or outperforms larger open-weight models (7B–32B) on most primary task metrics. We further deploy a LoRA-tuned Qwen3-VL-4B (Pocket-Dentist-4B) on a commodity smartphone, achieving 6.67s per-sample inference while keeping all computation local to the device.