siglip-base-patch16-224

Name: siglip-base-patch16-224
Author: google

by google

Sigmoid Loss for Language Image Pre-Training, efficient contrastive learning

1.4Mdl/month

89likes

203Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

google/siglip-base-patch16-224

Feature URI

mixpeek://image_extractor@v1/google_siglip_base_v1

Overview

SigLIP replaces CLIP's softmax-based contrastive loss with a simple pairwise sigmoid loss, enabling more efficient training on larger batch sizes without requiring a global normalization step.

On Mixpeek, SigLIP offers a lighter-weight alternative to CLIP for visual embedding extraction, with comparable accuracy on many benchmarks while being faster to run at inference time.

Architecture

Vision Transformer (ViT-B/16) with 12 layers, 768-dim hidden size, 12 attention heads. Uses sigmoid contrastive loss instead of softmax, eliminating the need for large batch normalization.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so siglip-base-patch16-224 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // The vector name has to match a vector index on the collection.
          vectors: { "image-embedding": yourVector },
          payload: { source_key: "archive/2026/asset-00412" },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// image_extractor@v1 runs google/siglip-base-patch16-224
// (768-d) over a bucket, with no inference of your own.

Capabilities

Efficient contrastive image-text learning
768-dimensional dense vector embeddings
Lower memory footprint than CLIP ViT-L
Strong zero-shot classification performance

Use Cases on Mixpeek

High-throughput visual indexing of large image catalogs

Real-time visual similarity for recommendation engines

Lightweight embedding extraction for edge deployments

Benchmarks

Dataset	Metric	Score	Source
ImageNet zero-shot	Top-1 Accuracy	73.2%	Zhai et al., 2023: Table 1
COCO (image→text)	Recall@1	62.7%	Zhai et al., 2023: Table 3
ObjectNet	Top-1 Accuracy	59.1%	Zhai et al., 2023: Table 2