EVA02-CLIP-L-14-336

Name: EVA02-CLIP-L-14-336
Author: BAAI

by BAAI

Enhanced CLIP visual encoder with masked image modeling pre-training at 336px resolution

19Kdl/month

430Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

BAAI/EVA02-CLIP-L-14-336

Feature URI

mixpeek://image_extractor@v1/baai_eva02_clip_large_v1

Overview

EVA02-CLIP-L-14-336 is a Vision Transformer CLIP model pre-trained with masked image modeling (MIM) to reconstruct language-aligned vision features, then fine-tuned with contrastive image-text learning. At 336px resolution with ~430M parameters, it achieves 80.4% zero-shot top-1 accuracy on ImageNet while using only ~1/6 the parameters and training data of the previous largest open-source CLIP.

On Mixpeek, EVA02-CLIP provides high-quality visual embeddings with better efficiency than giant CLIP models, powering image and video frame search with strong zero-shot generalization across domains.

Architecture

EVA02 Vision Transformer (ViT-L/14) with 24 layers, pre-trained via masked image modeling with CLIP feature reconstruction targets. Contrastive image-text fine-tuning on 6B image-text pairs. 336x336 pixel input resolution with patch size 14.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so EVA02-CLIP-L-14-336 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

80.4% zero-shot ImageNet top-1 (best in class for L-scale)
MIM pre-training for robust visual features
768-dimensional dense vector embeddings
336px high-resolution input for fine-grained details
1/6 parameters of comparable giant CLIP models

Use Cases on Mixpeek

High-quality visual search across image and video collections

Zero-shot classification of visual content without fine-tuning

Visual embedding extraction where accuracy matters more than speed

Benchmarks

Dataset	Metric	Score	Source
ImageNet zero-shot	Top-1 Accuracy	80.4%	Fang et al., 2023: EVA-CLIP paper
ImageNet fine-tuned	Top-1 Accuracy	90.0%	Fang et al., 2023: EVA-02 paper
ObjectNet	Top-1 Accuracy	72.3%	Fang et al., 2023: EVA-CLIP paper