Jina CLIP v2

jina-clip-v2 is Jina AI's general-purpose multilingual multimodal embedding model: a contrastive vision-language model (865M parameters) that embeds text and images into one shared space for cross-modal retrieval. Built on jina-embeddings-v3 as the text tower, it supports 89 languages, 8K-token text context, 512x512 image input, and Matryoshka dimensionality (64–1024) so you can trade embedding size for storage. It improved on v1 by ~3% across retrieval tasks and performs on par with frontier sub-1B multilingual embedders on MTEB while handling visually rich documents better than classic CLIP.

The honest picture: at 865M parameters it's ~2–3x slower per embed than smaller CLIP variants, so high-throughput pipelines need GPU batching (Infinity-Embedding is the recommended server). Multilingual claims are strong but uneven across low-resource languages. It's an embedding model, not a generator — a component for RAG and search, not a standalone product.