jina-embeddings-v5-omni-small
Jina AI
Jina AI's multimodal embedding model, released 7 May 2026. A ~1.74B-parameter model that encodes text, images, video, audio and PDF pages into a single shared 1024-dimensional vector space aligned with jina-embeddings-v5-text-small, so a corpus indexed as text can be queried with any modality without reindexing. It was trained by freezing the v5-text-small backbone and its four task LoRA adapters and training only the cross-modal projectors over a SigLIP2 So400m vision encoder and a Whisper-large-v3 audio encoder, so text-only output is bit-identical to v5-text-small. Max sequence length 32,768; Matryoshka truncation down to 32 dimensions. Weights are CC-BY-NC-4.0.
Strengths
- Encodes text, image, video, audio and PDF into one shared 1024-dimensional space, so index-in-one-modality / query-in-another needs no reindexing
- Text-only embeddings are bit-identical to jina-embeddings-v5-text-small, making it a drop-in upgrade with no reindexing of existing text corpora
- Four task-specific LoRA adapters carried over from v5-text-small: retrieval, text-matching, clustering, classification
- Matryoshka truncation to 768, 512, 256, 128, 64 or 32 dimensions for storage and latency trade-offs
- 32,768-token maximum sequence length for long documents
- Deployment breadth: Jina API, Elastic Inference Service, AWS SageMaker, Hugging Face and air-gapped installs, plus GGUF quantizations and MLX for Apple Silicon
- Video handled as 32 uniformly sampled frames and PDF pages rendered through the vision pathway, so document-page retrieval works without a separate OCR stage
Caveats
- Weights are CC-BY-NC-4.0 - non-commercial only; commercial use requires the hosted API, a CSP listing or a separate arrangement with Jina AI
- No multimodal benchmark scores are published as text on the vendor's model page or model card - the image/video/audio results appear only as a chart caption pointing at Table 1 of the arXiv report, so this record ships no benchmarks
- Text quality is unchanged from jina-embeddings-v5-text-small by construction (the backbone and adapters are frozen), so there is no text-retrieval gain over the smaller 677M model despite ~2.6x the parameters
- Vendor pages disagree on language coverage: the model page lists 32 trained / 93 supported languages while the v5-text HF card claims "119+ languages" for the shared text tower
- Vendor-published geometry diagnostics for the shared text tower are unflattering: effective dimensionality is given as 70 of 1024 with a 0.300 noise floor, and its retrieval "recall cliff" as 0.566
- Vendor parameter counts are inconsistent across its own surfaces - 1.7B in Model Details, ~1.74B in the overview and HF card, 1.6B in the Pareto-front annotation
- Vendor
- Jina AI
- Context Window
- 32,768 tokens
- License
- CC-BY-NC-4.0
- Release Date
- 2026-05-07
- Modalities
- textimageaudiovideo
Capabilities
Resources
Pricing
Weights are published under CC-BY-NC-4.0 (non-commercial). Hosted access is the Jina Search Foundation API (https://api.jina.ai/v1/embeddings), sold as prepaid token top-ups ("Top-up tokens as you go"); Jina's rate-limit table lists the Embedding API as blocked without a key, 100 RPM & 100,000 TPM on a free key, 500 RPM & 2,000,000 TPM on a paid key and 5,000 RPM & 50,000,000 TPM on a premium key. Also available via Elastic Inference Service, AWS SageMaker, Hugging Face and air-gapped deployment. Jina publishes no per-token price as page text, so no rate is asserted here.