siglip

Presents VectraYX-Vision-1B, a sub-2B Spanish/LATAM cybersecurity vision-language model coupling a frozen SigLIP encoder with a Spanish decoder via an MLP, yet reports near-zero visual grounding despite functional pipelines, with open-source weights and remediation plans.

A researcher reports a surprising 50-point accuracy gap between frozen SigLIP2 (92%) and DINOv2 (41%) embeddings on a fine-grained car classification task using k-NN, seeking insight on whether a linear probe would close the gap or if DINOv2 is unsuited for retrieval.