Quick Start

Get your first /v1/embeddings response in five minutes.

We'll use a prebuilt CPU image of jina-embeddings-v5-text-nano. No GPU needed, no Python, no bundle phase. Pure Docker.

The five-minute path

flowchart LR
    A[Pull image
from GHCR] --> B[Run container
docker run]
    B --> C[Health check
curl /health]
    C --> D[First embedding
curl /v1/embeddings]

Loading

Prerequisites

• Docker installed (docker ps returns without error)
• ~3GB free disk
• Port 8080 available
• A GitHub personal access token with read:packages scope (create one)

Air-gapped target host? Run steps 1-3 on a connected machine, transfer the .tar.gz, resume from step 4 on the offline host.

1. Authenticate to GHCR

Prebuilt images live in ghcr.io/jina-ai/jina-on-prem/*. They require auth (one-time):

echo "YOUR_GH_TOKEN" | docker login ghcr.io -u YOUR_GH_USERNAME --password-stdin

2. Pull the image

docker pull ghcr.io/jina-ai/jina-on-prem/jina-embeddings-v5-text-nano:cpu

~800 MB compressed, 2.9 GB on disk. Takes 1-3 minutes on a typical link.

3. (Optional) Export for offline transport

If the target host has no network:

docker save ghcr.io/jina-ai/jina-on-prem/jina-embeddings-v5-text-nano:cpu \
  | gzip > jina-v5-nano.tar.gz
# Transfer jina-v5-nano.tar.gz to the air-gapped host (SCP, USB, ...), then on the host:
docker load < jina-v5-nano.tar.gz

The repo ships scripts/pull-prebuilt.sh which does steps 1-3 in one command:

./scripts/pull-prebuilt.sh jina-embeddings-v5-text-nano cpu

4. Run

docker run -d --name jina-nano -p 8080:8080 \
  ghcr.io/jina-ai/jina-on-prem/jina-embeddings-v5-text-nano:cpu
docker logs -f jina-nano   # watch for "Uvicorn running on http://0.0.0.0:8080"

For GPU: add --gpus all and use the :gpu tag.

5. Verify

curl http://localhost:8080/health

{"status":"ok","model":"jinaai/jina-embeddings-v5-text-nano","device":"cpu","ready":true,"multimodal":false,"schemas":["openai","voyage","gemini","cohere"]}

First embedding:

curl -s http://localhost:8080/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"input":["Hello world"]}' \
  | jq '.data[0].embedding | length'
# 768

If both calls succeed, you're live. The same container also speaks Cohere, Gemini, and Voyage protocols - see API Reference for examples.

What you can do now

flowchart TB
    OK[Working /v1/embeddings] --> A[Wire into your app
OpenAI SDK with base_url=...]
    OK --> B[Hook up Elasticsearch
inference service]
    OK --> C[Try a bigger model
v5-text-small or v5-omni]
    OK --> D[Add a reranker
jina-reranker-v3]
    OK --> E[Bundle your own
Bundling Guide]

Loading

• Customer Scenarios - end-to-end deployment patterns
• API Reference - all four schemas, multimodal, ES integration
• Picking a Model - decision tree for the other 27 models
• Bundling Guide - build your own bundle for models not in the prebuilt list

When things go wrong

The three errors that catch most first-time users:

• permission denied while trying to connect to the docker API -> you're not in the docker group. Fix.
• Error response from daemon: error from registry: unauthorized -> not logged into GHCR. Run step 1.
• bind: address already in use -> something's on port 8080. Map a different port: -p 9090:8080.

Full list: Troubleshooting.