clip-vit-base-patch32

Name: clip-vit-base-patch32
Author: openai

by openai

The most-downloaded CLIP variant: fast zero-shot image and text embeddings

20.0Mdl/month

1,005likes

151Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

openai/clip-vit-base-patch32

Feature URI

Overview

CLIP ViT-B/32 is the small, fast member of the original CLIP family, and by download volume it is the most widely used vision model on Hugging Face. It maps images and text into one shared embedding space, so a text query can retrieve images with no labelled training data.

The patch32 in the name is why it is fast: it divides a 224x224 image into 32x32 pixel patches, giving 49 patches per image instead of the 256 that ViT-L/14 produces. Fewer patches means much less compute per image, at some cost in fine detail. For large media libraries where you embed millions of frames, that is usually the right trade, which is why this variant dominates production retrieval workloads.

On Mixpeek it backs image and video-frame embedding, turning frames into vectors you query with natural language through one API.

Architecture

Vision Transformer (ViT-B/32): 12 layers, 768-dim hidden size, 32x32 pixel patches, 49 patches per 224x224 image. The text encoder is a 12-layer transformer. Both project into a shared 512-dim embedding space trained with a contrastive objective over 400M image-text pairs.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so clip-vit-base-patch32 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.