Picking A Model

There are 28 models. Most customers need one of six. Use this decision tree.

flowchart TD
    A[What is the customer doing?] --> B{Just text?}
    B -- yes --> C{Multilingual?}
    B -- no --> D{Modalities?}
    C -- mostly English, latency matters --> M1[jina-embeddings-v5-text-nano
239M, ~2GB VRAM]
    C -- 89 languages, quality matters --> M2[jina-embeddings-v5-text-small
677M, ~3GB VRAM]
    C -- need 32K context --> M3[jina-embeddings-v5-text-small
or v4 for biggest]
    D -- text + image --> M4[jina-clip-v2
865M, ~4GB VRAM]
    D -- text + image + audio + video --> M5[jina-embeddings-v5-omni-small
1.74B, ~8GB VRAM]
    D -- code --> M6[jina-code-embeddings-1.5b
1.5B, ~4GB VRAM]

    A --> E{Reranking?}
    E -- top quality --> R1[jina-reranker-v3
597M, ~3GB VRAM, 131K context]
    E -- fastest --> R2[jina-reranker-v1-turbo-en
37.8M, CPU OK]

    A --> F{HTML/document cleanup?}
    F -- yes --> RD1[ReaderLM-v2
1.54B, 512K context]

Loading

Quick reference by use case

Use case	Model	Why
First demo / proof of concept	jina-embeddings-v5-text-nano	smallest, fastest to bundle, CPU works
Production multilingual semantic search	jina-embeddings-v5-text-small	89 languages, task-aware, 32K context
Drop-in for OpenAI text-embedding-3-small	jina-embeddings-v5-text-nano	similar dim (768), credible quality
Multimodal RAG (PDFs with images)	jina-embeddings-v4 or v5-omni-small	image + text in same space
Image search	jina-clip-v2	optimized for text + image
Code search	jina-code-embeddings-1.5b or 0.5b	trained on code
Reranking	jina-reranker-v3	best quality, long context
Cheap fast reranking	jina-reranker-v1-turbo-en	38M params, CPU OK
HTML to clean markdown	ReaderLM-v2	512K context, document focused
Legacy English embeddings	jina-embeddings-v2-base-en	Apache-2.0, no commercial license needed

Trade-offs

Bigger isn't always better. v5-text-small is a stronger model than v5-text-nano, but the nano is 3x smaller and runs on CPU. For a customer with tight latency SLA and English-only data, nano often wins.

CC-BY-NC-4.0 license matters for sales. All v5/v4/v3 models are CC-BY-NC-4.0 - the customer needs a commercial license from Elastic. v2 and v1 models are Apache-2.0 and free for any use. If the customer is allergic to vendor lock-in or wants to evaluate without buying, start with jina-embeddings-v2-base-en (or -zh, -de, -es).

Multimodal models cost more memory. A v5-omni-small needs ~8GB VRAM. A v5-text-nano needs 2GB. If the customer doesn't actually need image/audio/video, don't pay for it.

Matryoshka truncation lets you downsize at query time. v5/v4/v3 support requesting any dim from 32 to the model's max (768 / 1024 / 2048). A customer can index at 1024-dim and query at 128-dim if they want a smaller index. Pass dimensions: 128 in the request.

When to use multiple models

A real production stack often combines:

flowchart LR
    Q[User query] --> EMB[Embed query
v5-text-small]
    EMB --> VS[(Vector store
ES with v5-small index)]
    VS -->|top 100| RR[Rerank
reranker-v3]
    RR -->|top 10| APP[App]

Loading

• Embedding model for indexing and first-pass retrieval
• Reranker for high-precision top-K
• Optional ReaderLM if input is messy HTML/PDF

Both models can run side-by-side in two containers, OR you can deploy two replicas of each behind a load balancer if QPS demands it. See Sizing & Hardware for capacity planning.

Next

• Sizing & Hardware - how much GPU/RAM/disk
• Model Catalog - full specs for all 28
• Customer Scenarios - applied examples