PaliGemma2 3B

PaliGemma2 3B is Google's compact vision-language model, announced December 5, 2024, as part of the PaliGemma 2 family in 3B, 10B, and 28B sizes. It pairs the SigLIP image encoder with the Gemma 2 2B language model, with the extra vision parameters bringing the total to roughly 3B.

The family introduced flexible input resolutions of 224, 448, and 896 pixels and a simplified fine-tuning story, positioned as a drop-in replacement for the original PaliGemma. Google highlights long-form image captioning, chemical formula recognition, music score interpretation, spatial reasoning, and chest X-ray report generation among its strengths.

Weights are distributed under the Gemma license, which permits redistribution, commercial use, fine-tuning, and derivatives, though the custom terms should be reviewed before commercial deployment. Checkpoints ship in bfloat16 with transformers, Keras, PyTorch, and JAX support, plus Kaggle and Ollama availability.