
Jina AI, part of Elastic, released jina-embeddings-v5-omni-nano on 2026-05-07. It is a 1.04B parameter multimodal embedding model that maps text, images, video, audio and PDF pages into one shared vector space, with a 768-dimension output and an 8K token context. Its text embeddings are identical to jina-embeddings-v5-text-nano, so teams can add media to an existing text index without reindexing. Weights are open on Hugging Face under CC-BY-NC-4.0, which requires a commercial license for commercial use.
A workable 1.04B-parameter 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.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
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 | 1.2 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 1.2 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 1.2 GB |
| AMD Instinct MI300XAMD | SS | 1.2 GB |
| AMD Instinct MI325XAMD | SS | 1.2 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 Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.07 |
NVIDIA Tesla V100 16GBVast.ai · On-Demand · 16 GB VRAM | $0.07 |
NVIDIA GeForce RTX 4060 TiVast.ai · Spot · 8 GB VRAM | $0.08 |
NVIDIA GeForce RTX 3080Vast.ai · Spot · 10 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, operating as part of Elastic, released jina-embeddings-v5-omni-nano on May 7, 2026. Built as a compact 1.04B parameter dense embedding model, it tackles one of the most frustrating bottlenecks in multimodal information retrieval: vector space fragmentation. Most embedding pipelines require separate vector spaces for images, audio, and text, forcing engineering teams to maintain disparate indices or recalculate historical data when upgrading pipelines.
jina-embeddings-v5-omni-nano maps text, images, video, audio, and PDF documents into a single, unified 768-dimensional vector space. Crucially, its text representations are mathematically identical to jina-embeddings-v5-text-nano at FP16 precision. If your team already has an established vector database populated with text entries from Jina's text-nano model, you can ingest image, video, and audio assets directly into that existing index without reindexing a single text document.
Positioned on the Pareto efficiency frontier for open-weight multimodal embeddings, this 1.04B model targets edge deployments, local workstations, and air-gapped enterprise search engines. It trades the massive footprint of 7B-plus multimodal encoders for an ultra-compact footprint that fits comfortably inside consumer VRAM, delivering low-latency feature extraction without sacrificing cross-modal retrieval fidelity.
The core architecture of jina-embeddings-v5-omni-nano is a 1.04B parameter dense model that uses locked aligned towers. Instead of training all modal encoders jointly from scratch and destabilizing text geometry, Jina locks the underlying text backbone to match jina-embeddings-v5-text-nano. Visual, audio, and document encoders (including Whisper-derived audio blocks and vision encoders) are aligned directly against this locked target space.
Key technical specifications include:
.jpg, .png, .webp, .svg, .heic), video (.mp4, .mkv, .webm), audio (.wav, .mp3, .flac, .opus), and multi-page documents (.pdf)The model relies on built-in LoRA-style task adapters rather than requiring completely separate checkpoints for every workflow. The base repository ships with four specialized adapter modes:
retrieval: Asymmetric document and query search (prepends Query: or Document: formatting).classification: Optimized embedding boundaries for zero-shot and few-shot linear probing.clustering: Geometry tuned for semantic group discovery and deduplication.text-matching: Symmetric semantic similarity scoring (STS) and duplicate detection.The 8,192-token context length is particularly notable for a dense 1B model. It allows practitioners to encode long-form technical reports, audio transcripts, or continuous video frames into a single unified vector without aggressive chunking artifacts.
jina-embeddings-v5-omni-nano is built specifically for retrieval-augmented generation (RAG) pipelines and enterprise search over mixed-media asset stores.
Engineers can build search pipelines where a natural-language text query directly retrieves relevant video timestamps, podcast audio clips, slide decks, or technical PDF diagrams. Because the embedding space is shared, querying Find the circuit diagram for the step-down converter retrieves the exact schematic page from a PDF manual or an annotated frame inside a video tutorial.
Using the model's clustering adapter, digital asset management systems can cluster duplicate and near-duplicate assets across formats. An audio recording of an interview, its text transcript, and a video recording of the same session project to adjacent coordinates, enabling automatic grouping and deduplication across file formats.
By reducing dimensional output via Matryoshka representation learning (e.g., truncating from 768 down to 256 or 128 dimensions), you can run zero-shot classification pipelines locally on constrained edge hardware. This fits workflows like on-device photo sorting, audio event classification, and content moderation in bandwidth-limited environments.
Because this is a dense 1.04B parameter network, the hardware barrier to entry is minimal. You do not need an enterprise cluster or multiple data-center GPUs to run high-throughput local inference.
The memory footprint depends on the execution precision and batch size:
Any modern consumer card can run this model locally with headroom to spare:
For generative language models, 4-bit quantization (Q4_K_M) is often the sweet spot. However, for embedding models, representation geometry degrades faster under aggressive quantization.
Because jina-embeddings-v5-omni-nano VRAM requirements are already remarkably low (under 2.5 GB in unquantized FP16), the recommended deployment strategy is to run in FP16 or BF16 natively whenever possible. If you must run on constrained devices, such as an integrated GPU or mobile edge node, use INT8 or Q8_0. Reserve Q4_K_M strictly for embedded scenarios with severe RAM ceilings (sub-2GB system budgets), as 4-bit quantization can soften the subtle distance margins necessary for difficult cross-modal retrieval.
Practitioners have multiple pathways to run the model locally:
AutoModel with trust_remote_code=True or native sentence-transformers).jina-embeddings-v5-omni-nano-retrieval.In terms of local performance, jina-embeddings-v5-omni-nano tokens per second regularly exceeds 1,200 to 2,500 tokens per second for text processing on an RTX 4080 or RTX 4090 at batch sizes of 32 or 64. Audio and image encoding latencies vary based on resolution and duration, but typical image embedding passes take under 25 milliseconds per image on modern hardware.
Understanding where jina-embeddings-v5-omni-nano fits among alternative local embedding models helps clarify whether to deploy it over competing options.
LanguageBind was an early landmark for open multimodal representation, but its cross-modal alignment suffers on fine-grained document text and modern MIEB image benchmarks. On multi-media evaluations (including MAEB audio and MMEB video), jina-embeddings-v5-omni-nano outperforms LanguageBind across retrieval metrics by more than 11 points while operating at a slightly smaller parameter budget (1.04B vs. 1.14B). Jina also provides native 8K context and four task adapters, whereas LanguageBind is limited by short text context windows and lack of asymmetric search tuning.
Within Jina's own v5 lineup, omni-small steps up to 1.57B parameters and competes directly with much larger models like LCO-7B. Choosing between them comes down to deployment constraints:
Note that jina-embeddings-v5-omni-nano is published under a Creative Commons Attribution-NonCommercial 4.0 license (CC-BY-NC-4.0). While ideal for local prototyping, internal academic research, and non-commercial development, commercial production deployments require obtaining a commercial license through Jina AI or running managed inference through Elastic.

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.