PaliGemma

PaliGemma, released by Google DeepMind in May 2024, is the first vision-language model of the Gemma family: a ~3B-parameter model combining the SigLIP-So400m vision encoder with the Gemma 2B language decoder, pretrained on Google's WebLI dataset. It comes in 224px, 448px, and 896px resolutions and in pretrained and multi-task-mix variants.

Key nuance: PaliGemma is a transfer base designed to be fine-tuned for tasks like captioning, VQA, OCR, and segmentation — it is not a chat model and performs poorly at zero-shot conversation. It is distributed under the Gemma Terms of Use, which allow commercial use but are a custom license, not Apache 2.0 or another OSI-approved license.