Visual Salamandra 7B extends Salamandra to image and video inputs with SigLIP encoder

AI Impact Summary

Visual Salamandra extends the Salamandra 7B foundation model with a SigLIP encoder and a multilayer projector to enable image and video inputs via a late-fusion architecture. The four-phase training emphasizes vision-language alignment and multilingual instruction tuning, targeting European language diversity for robust multimodal understanding. This unlocks VQA, OCR, and document/graph understanding at a compact 7B footprint, but users should validate OCR on dense layouts and monitor for hallucinations in ambiguous visuals. The model is released under Apache License 2.0 for research and non-commercial use, implying licensing considerations for commercial deployments.