siglip
256M-param document VLM: SigLIP + SmolLM2. LoRA fine-tuning, ONNX export, runs on CPU. Apache 2.0.
Search long video without running a VLM on every frame. Index once with a small encoder, spend the expensive model only where it matters.
A minimal, but effective implementation of CLIP (Contrastive Language-Image Pretraining) in PyTorch
