SigLIP 2

SigLIP 2 is Google DeepMind's second-generation vision-language encoder, released in February 2025. It rebuilds the original SigLIP training recipe with sigmoid loss plus decoder-based pretraining and self-distillation with masked prediction, yielding better semantic understanding, localization, and dense visual features. It supports 109 languages and ships in sizes from 86M Base up to 1B Giant, with fixed-resolution and native-aspect-ratio variants.

All weights are Apache-2.0, making it one of the most commercially friendly vision encoders available. It is the default backbone for a large share of open multimodal pipelines, retrieval systems, and on-device similarity search, and it underpins other open models such as VideoLLaMA 3. Google claims state-of-the-art zero-shot classification, retrieval, and dense prediction in the paper.

Google claims state-of-the-art on zero-shot classification, image-text retrieval, dense prediction, and localization (paper arXiv:2502.14786); 109-language multilingual coverage (all provider claims)