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

← Catalog

google/siglip-so400m-patch14-384

Open comparison →

Primitive: /encode · Encode · SigLIP

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.

MultimodalDense

Overview

Hardware: — drives latency, throughput & cost

Size878M params
Tasks /encode
Licenseapache-2.0
Latency347 ms
Throughput451 tok/s
Cost$0.493 /1M tok

Cost is approximate — computed from list GPU prices; your actual price depends on the provider you deploy SIE with.

Embedding

Output typesDense
Dimensionsdense: 1,152
Max sequence length64
Inputstext · image

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 0.8365
mrr at 10 0.9663
Performance L4-SPOT b1 c8
Corpus 202 tok/s
Corpus p50 523.6ms
Query 9.7 img/s
Query p50 711.3ms
Performance L4 b1 c16
Corpus 700 tok/s
Corpus p50 170.9ms
Query 7.3 mpix/s
Query p50 197.2ms
Reference →

Open source inference for agents

Open-source inference for the models behind your agents. Run it yourself, or let us run it for you.

Github 2.1K

Contact us

Tell us about your use case and we'll get back to you shortly.

Apply for an inference grant

Free capacity on our hosted cluster for selected projects. Tell us what you run and we reply by email.