SmolVLM, Florence-2 top tiny VLM picks

Tiny VLMs are the final nail in the coffin for expensive, task-specific image classifiers. Nuanced moderation no longer requires a GPU or a massive foundation model. SmolVLM-256M and Florence-2-base provide 5-10 it/s throughput on standard processors, and their "no-refusal" descriptive capabilities make these models ideal for explicit content tagging and filtering. Quantization via ONNX Runtime or OpenVINO is essential for hitting performance targets on CPU, enabling real-time, nuanced visual reasoning at the edge for a fraction of the cost.