PaliGemma: A versatile 3B VLM for transfer

Task-specific VLM fine-tuning was fragmented across many bespoke models and datasets.

SigLIP-So400m vision encoder plus the Gemma-2B language model, trained across ~40 tasks with a mix of task prefixes. The composition is deliberately plain, and the vision tower is the same SigLIP encoder recorded in pap-mm-siglip.

Technical report; no peer-reviewed venue found. PaliGemma is shipped inside production stacks partly BECAUSE of its shape: a small open VLM with a frozen, standard vision tower, so its serving profile is predictable - one SigLIP pass at fixed resolution plus a 2B-parameter decoder, which fits comfortably on a single L4 or L40S. PaliGemma 2 added native resolution and Pan and Scan cropping for high-resolution inputs, which reintroduces the multi-crop token cost. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.