ollama-export

Export Ollama models to tar archives or push them directly to OCI-compliant registries. Built for airgapped and private-network environments where you cannot reach the public Ollama registry.

Features

• Pull models from the Ollama registry via the local Ollama daemon
• Export any locally cached model to a self-contained tar archive
• Import a tar archive back into the local Ollama model store
• Push directly to an OCI registry (Harbor, Nexus, Docker Registry v2, …)
  • Chunked PATCH uploads — works correctly behind Traefik and other proxies
  • Automatic CSRF token handling (Harbor)
  • Bearer token + HTTP Basic auth with challenge negotiation
  • Blob deduplication (skips blobs already present on the target)

Installation

From PyPI (recommended)

pip install ollama-exporter

From source

git clone https://github.com/thenomadbeyond/ollama-export
cd ollama-exporter
pip install -e .

Standalone binary (Linux / macOS)

Pre-built binaries are available on the Releases page — no Python required.

chmod +x ollama-exporter-linux-amd64
./ollama-exporter-linux-amd64 --help

Requirements

• Python 3.9+
• A running ollama serve instance (default: http://localhost:11434)

Usage

List local models

ollama-exporter list

Pull a model from the Ollama registry

# Public model
ollama-exporter pull llama3.1:8b

# Private registry — inline credentials
ollama-exporter pull registry.example.com/myorg/mymodel:v1 -u alice -p secret

# Private registry — password from stdin (avoids credentials in shell history)
echo "$TOKEN" | ollama-exporter pull registry.example.com/myorg/mymodel:v1 \
    -u alice --password-stdin

# Private registry — credentials from environment variables
export OLLAMA_USERNAME=alice
export OLLAMA_PASSWORD=secret
ollama-exporter pull registry.example.com/myorg/mymodel:v1

Export to a tar archive

ollama-exporter export mistral:latest mistral-latest.tar
ollama-exporter export llama3.1:8b /mnt/usb/llama3.1-8b.tar

Import from a tar archive

# Model name is inferred from the filename
ollama-exporter import mistral-latest.tar

# Override the name/tag on import
ollama-exporter import /mnt/usb/llama3.1-8b.tar --tag llama3.1:8b

Push to a private OCI registry

# Inline credentials
ollama-exporter push mistral:latest registry.example.com/ai/mistral:latest \
    -u admin -p secret

# Password from stdin (CI-friendly, nothing in shell history)
echo "$TOKEN" | ollama-exporter push mistral:latest registry.example.com/ai/mistral:latest \
    -u admin --password-stdin

# Credentials from environment variables
export OCI_USERNAME=admin
export OCI_PASSWORD=secret
ollama-exporter push mistral:latest registry.example.com/ai/mistral:latest

# Plain HTTP registry (e.g. internal dev)
ollama-exporter push qwen2.5-coder:7b registry.local/ai/qwen2.5-coder:7b \
    --insecure

# Tune chunk size if the proxy has tight timeouts (default: 64 MiB)
ollama-exporter push llama3.1:8b registry.example.com/ai/llama3.1:8b \
    --chunk-size 16

Airgapped workflow

[internet machine]
  ollama-exporter pull llama3.1:8b
  ollama-exporter export llama3.1:8b llama3.1-8b.tar
  # copy tar to USB drive / sftp / object storage

[airgapped machine]
  ollama-exporter import llama3.1-8b.tar
  # model is immediately available to ollama serve

Or, if your private network has an OCI registry reachable from both sides:

[internet machine]
  ollama-exporter pull llama3.1:8b
  ollama-exporter push llama3.1:8b registry.internal/ai/llama3.1:8b

[airgapped machine]
  ollama pull registry.internal/ai/llama3.1:8b

Environment variables

Variable	Default	Description
OLLAMA_HOST	http://localhost:11434	Ollama API base URL
OLLAMA_MODELS	~/.ollama/models	Ollama model store path
OLLAMA_USERNAME	—	Username for pull from a private registry
OLLAMA_PASSWORD	—	Password for pull from a private registry
OCI_USERNAME	—	Username for push to an OCI registry
OCI_PASSWORD	—	Password for push to an OCI registry

Registry compatibility

Registry	Tested
Harbor	Yes
Docker Registry v2	Yes
Nexus OSS	Should work
Gitea / Forgejo	Should work
AWS ECR	Untested
GHCR	Untested

How it works

Ollama stores models on disk in OCI format:

~/.ollama/models/
├── manifests/registry.ollama.ai/library/<name>/<tag>   # OCI manifest JSON
└── blobs/sha256-<hash>                                  # binary blobs

ollama-exporter reads these files directly — no re-downloading from the internet. A tar export packages the manifest and all referenced blobs. An OCI push uploads each blob via the OCI Distribution Spec v2 API and then uploads the manifest.

Security

See SECURITY.md.

License

Apache 2.0