Why did we open-source our inference engine? Read the post

google/siglip-so400m-patch14-384

SigLIP model pre-trained on WebLi at resolution 384x384. It was introduced in the paper Sigmoid Loss for Language Image Pre-Training by Zhai et al. and first released in this repository.

Architecture
SigLIP
Parameters
878M
Tasks
Encode
Outputs
Dense
Dimensions
Dense: 1,152
Max Sequence Length
64 tokens
License
apache-2.0

Benchmarks

Flickr30kI2TRetrieval

general retrieval en

Image-to-text retrieval: retrieve captions from images

Corpus: 31,783 Queries: 1,000
Quality
ndcg at 10 0.9001
map at 10 0.8364
mrr at 10 0.9663
Performance L4-SPOT b1 c8
Corpus TPS 202
Corpus p50 523.6ms
Query TPS 10
Query p50 711.3ms
Performance L4 b1 c16
Corpus TPS 508
Corpus p50 452.9ms
Query TPS 18
Query p50 551.4ms
Reference →

Self-hosted inference for search & document processing

Cut API costs by 50x, boost quality with 85+ SOTA models, and keep your data in your own cloud.