Jina CLIP v2
Jina CLIP v2 aligns text and images in one embedding space. It combines a 561-million-parameter Jina-XLM-RoBERTa text tower with a 304-million-parameter EVA02-L14 vision tower.
How the model works
The text encoder accepts up to 8,192 tokens, while the image encoder uses 512-by-512 inputs. Joint training supports text-to-image and text-to-text retrieval. Both towers return 1,024-dimensional representations that can be shortened to 64 dimensions. The model card lists 89 languages for multilingual image retrieval.Jina CLIP v2 model details
Use and access
The larger image input distinguishes v2 from v1's 224-by-224 input. Its text tower uses mean pooling, while its image tower uses the classification token. Jina provides weights and an embedding API. CC BY-NC 4.0 applies to the v2 weights, unlike the Apache 2.0 license on v1. CLIP v2 retrieves or matches content; it does not generate pictures.Jina CLIP v2 model detailsJina CLIP v2 technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | jina-clip-v2jinaai/jina-clip-v2 |
| Context window | 8,192 tokens Maximum input sequence in the model card; serving limits can be lower. Jina CLIP v2 model details |
| Parameters | 865 million across text and vision towersJina CLIP v2 model details |
| Embedding dimensions | 1,024; reducible to 64Jina CLIP v2 model details |
| Weight license | CC BY-NC 4.0Jina CLIP v2 model details |
| Base model | Jina-XLM-RoBERTa and EVA02-L14Jina CLIP v2 model details |
Related models
- Jina Embeddings v5 Omni Nano
Jina AI · Jina Embeddings v5 Omni
A compact multimodal encoder aligned with Jina's v5 Text Nano vectors.
Launch date API · Downloadable - Jina CLIP v1
Jina AI · Jina CLIP
An English embedding model trained for both text search and image matching.
Launch date Downloadable