Jina AI, part of Elastic, released jina-embeddings-v5-omni on 2026-05-11, a multimodal embedding family that maps text, images, audio and video into one vector space. It ships in two sizes, small at 1.57B parameters and nano at 0.95B, and keeps the frozen jina-embeddings-v5-text backbone so existing text indexes do not need rebuilding. The small model averages 54.04 across the four modalities and scores 67.0 on MMTEB text, with support for more than 100 languages. Elastic added the models to the Microsoft Foundry catalog on 2026-11-18.
A workable dense embedding model from Jina AI. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | SS | 0.5 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 0.5 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 0.5 GB |
| AMD Instinct MI300XAMD | SS | 0.5 GB |
| AMD Instinct MI325XAMD | SS | 0.5 GB |
Cheapest current cloud rentals with at least 1 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 3060Vast.ai · Spot · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 3060Vast.ai · On-Demand · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 5070Vast.ai · Spot · 12 GB VRAM | $0.07 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.08 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
Jina AI released jina-embeddings-v5-omni to solve one of the most frustrating bottlenecks in multimodal search: cross-modal space alignment without forcing full vector database reindexing. Built as a family of dense embedding models, jina-embeddings-v5-omni maps text, images, audio, and video into a single shared 1024-dimensional vector space. It comes in two primary configurations: a small variant at 1.57B parameters and an ultra-compact nano variant at 0.95B parameters.
The central design choice of this release is backward compatibility. The architecture locks the text backbone from jina-embeddings-v5-text-small, ensuring that existing text vector indexes remain completely valid. Practitioners do not need to re-encode millions of text documents to begin querying them with audio clips, images, or video snippets. In synthetic and real-world retrieval benchmarks, the 1.57B small model delivers an average score of 54.04 across all four modalities while reaching 67.0 on MMTEB text tasks, performing within half a point of models five times its size.
For developers seeking to run jina-embeddings-v5-omni locally, the modest parameter footprint makes deployment straightforward on consumer-grade hardware. Rather than relying on cloud APIs, engineers can host the dense encoders on entry-level GPUs or Apple Silicon unified memory systems, maintaining complete data privacy for proprietary media archives and local retrieval-augmented generation (RAG) pipelines.
The core mechanism powering jina-embeddings-v5-omni relies on locked aligned towers. Instead of training all modal encoders jointly from scratch, which often degrades text retrieval performance through modality collapse, Jina AI froze the text encoder and trained the vision and audio projectors against this anchor. This preserves the geometric distribution of the text space while pulling visual and acoustic features into alignment.
Key technical specifications across the model family include:
The integration of Matryoshka embeddings provides a massive operational advantage for local and edge search setups. You can truncate vector outputs to 256 or 512 dimensions to reduce vector index memory footprint and accelerate cosine distance calculations in engines like Qdrant, Milvus, or FAISS, with negligible degradation in ranking quality. Furthermore, the model family ships with four distinct task-specific heads: retrieval (asymmetric query-document search), classification (zero-shot label mapping), clustering (topic discovery and deduplication), and text-matching (symmetric semantic textual similarity).
Rather than treating multimodal search as an ensemble of disjoint encoders, jina-embeddings-v5-omni projects multiple data types into a unified vector coordinate system. This enables practical cross-modal pipelines:
Because the model occupies the sub-2B parameter space, the hardware barrier to entry is minimal. Local practitioners do not need multi-GPU server clusters; a standard modern laptop or budget desktop card handles inference comfortably.
The small variant (1.57B parameters) exhibits a very manageable memory profile:
For macOS users, Apple Silicon systems (M2, M3, or M4 with unified memory) run the model natively via MLX or GGUF ports. A baseline Mac mini with 16 GB of RAM can easily host the model alongside a vector database and an LLM without swapping memory.
When determining the best quantization for jina-embeddings-v5-omni, the operational trade-off differs from standard generative text models. In generative tasks, 4-bit quantization occasionally introduces syntactic degradation; in dense embedding models, aggressive quantization can distort high-dimensional cosine similarity distances.
For production semantic search, FP16 or Q8_0 is the recommended format whenever hardware allows. Since FP16 requires less than 4 GB of VRAM, any GPU with 6 GB or more should run unquantized weights. If you are deploying on resource-constrained devices, such as a Raspberry Pi 5 or low-power edge gateways, Q4_K_M delivers acceptable retrieval recall while keeping memory consumption under 1.5 GB.
On an NVIDIA RTX 4090, throughput for the 1.57B small model reaches hundreds of texts per second in batch mode, processing typical text documents at over 2,500 tokens per second. For media inputs, single-image or audio-clip feature extraction typically completes in 15 to 40 milliseconds per item depending on patch resolution and audio length.
The model can be run via Hugging Face transformers or sentence-transformers, with community GGUF weights supported across llama.cpp-compatible tooling:
1# Basic setup using sentence-transformers2pip install sentence-transformers einops timm
1from sentence_transformers import SentenceTransformer23# Load the base omni model with retrieval adapter4model = SentenceTransformer(5 "jinaai/jina-embeddings-v5-omni-small",6 trust_remote_code=True7)89# Encode text and multimodal assets into the same 1024-d space10embeddings = model.encode([11 "Query: architectural blueprints of industrial cooling towers",12 "Document: Cooling tower operational schematics and pipe dimensions."13])
For users who want to run jina-embeddings-v5-omni locally through command-line runners, quantized GGUF binaries can be loaded into local inference runtimes or imported into Ollama once custom Modelfiles for the multi-tower vision-audio layers are configured.
To understand where jina-embeddings-v5-omni fits, it is best compared against existing open-weight omni and multimodal embedding baselines.
LanguageBind was one of the earliest models to project multiple modalities into a joint representation space. However, LanguageBind suffers from severe text-retrieval degradation compared to modern dedicated text embedding models. The 0.95B nano variant of jina-embeddings-v5-omni outperforms LanguageBind by more than 11 points on unified benchmarks, while the 1.57B small variant dramatically outclasses LanguageBind on MMTEB text benchmarks (67.0 vs. sub-50 scores). For enterprise search where text remains the primary query vehicle, jina-embeddings-v5-omni is substantially superior.
The LCO series represents the high-end frontier of open multimodal embedding models. While LCO-7B achieves an overall benchmark average of 54.43, it requires an 8.93B parameter execution footprint, requiring at least 18 GB of VRAM in unquantized mode. By contrast, jina-embeddings-v5-omni-small achieves a 54.04 average across text, image, audio, and video at only 1.57B parameters. That delivers 99% of the quality at a 5.7x parameter reduction, allowing developers to run unified multimodal retrieval on consumer hardware rather than dedicating high-tier data center GPUs to vector indexing.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Jina AI model we track.

Explore the Family
The full Jina Embeddings family leaderboard with sizes, benchmark scores, and a release timeline.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.