Assemble and Propagate AI artifacts Containerization on Kubernetes - APACK

apack is a cloud-native AI model packaging and distribution tool designed for the modern AI/ML lifecycle. It containerizes AI models, making them first-class citizens alongside Docker containers, fundamentally transforming how models are delivered, distributed, and executed.

Core Value: Manage AI models like Docker images. Zero learning curve, out-of-the-box experience, enabling standardized version control, build packaging, debugging, and delivery. Build once, run anywhere.

Core Features

100% Cloud-Native Architecture

• OCI Standard Compliance: Fully compliant with Open Container Initiative (OCI) standards - model images are container images.
• Universal Runtime: Seamlessly compatible with mainstream container runtimes like docker, containerd, cri-o, supporting direct model pulling and mounting.

Minimalist Operation Experience

• No Dockerfile Required: Eliminate complex build configurations. The tool handles underlying details automatically - no need to be a container expert.
• Built-in Inference Engine: Out-of-the-box functionality, no complex configuration required. One-click local model execution with built-in (llama.cpp/llamafile) runtime and plugin support for other mainstream frameworks (PyTorch/TensorFlow/ONNX).

Cost Optimization & Efficiency

• Smart Layered Packaging: Leverage OCI's layered storage mechanism, focusing only on changed files and layers (e.g., fine-tuned model weights only).
• Incremental Updates: Achieve zero redundancy at node-level model resources, significantly reducing storage footprint and network bandwidth.

Security & Compliance

• Full-Chain Packaging: Package not only model weights but also datasets, inference code, dependencies, and documentation, forming complete product-level AI/ML data chains.
• Offline Delivery: Support secure delivery in private, offline environments.
• Audit-Friendly: Meet data and model traceability and compliance requirements through image version management.

Comparative Advantages

Feature	Traditional Model Deployment	apack (Model Containerization)
Distribution	File transfer (tar/zip) / HuggingFace CLI	OCI image pull (Docker/Containerd)
Environment Dependencies	Manual configuration, prone to inconsistencies	Environment solidify within images, 100% consistency
Update Mechanism	Full download/overwrite	Incremental updates, only transfer differential layers
Local Execution	Requires Python environment setup	apack run direct execution, built-in engine
Cloud Production	Dedicated CLI or scripts	Reuse Docker CLI / Kubernetes

Workflow

apack starts from trained model files, analyzes and generates a declarative description file (Apackfile) in the given model directory. This file serves as the build instruction for model images and is stored alongside model files. Users specify this file when building images to create model containers. Images are distributed under oci-image-spec:v1.1, compatible with all image registries and cloud-native container tools like docker and containerd.

1. Built-in inference engine enables local execution and debugging without dependency on docker, containerd, or other tools. Zero configuration - the tool is ready to use immediately after download.
2. Locally debugged model images can be directly pushed to registries and deployed to the cloud without secondary packaging or intermediate processing.

Note

Cloud deployment uses Kubernetes ImageVolume feature. v1.33/v1.34 require manual enabling of this feature, while v1.35 enables it by default.

Quick Start

System Requirements

Local Environment

localization-mode.mp4

Component	Minimum Requirements	Recommended Configuration
Operating System	Linux/macOS/Windows	Linux Ubuntu 20.04+
Memory	4GB RAM	16GB+ RAM
Storage	10GB free space	50GB+ SSD
Network	Stable internet connection	Bandwidth ≥ 10Mbps

Cloud Environment

kube-native.mp4

Component	Minimum Requirements	Recommended Configuration
Container Runtime	Docker 20.10+ or Containerd 1.6+	Docker 24.0+
Kubernetes	v1.33	v1.35
Operating System	Kernel 4.5	Kernel 5.1+
Memory	128GB RAM	512GB+ RAM
Storage	500GB free space	5TB+ SSD
Network	Enterprise-grade connection	Bandwidth ≥ 100Mbps
Image Registry	OCI-compliant registry	Harbor/AWS ECR/Azure ACR

Install apack

Method 1: Pre-compiled Binaries (Recommended)

Linux

# AMD64 Architecture
curl -L https://github.com/model-ci/apack/releases/latest/download/apack-linux-amd64 -o apack
https://github.com/model-ci/apack/releases/tag/0.0.1
chmod +x apack
sudo mv apack /usr/local/bin/

# ARM64 Architecture
curl -L https://github.com/model-ci/apack/releases/latest/download/apack-linux-arm64 -o apack
chmod +x apack
sudo mv apack /usr/local/bin/

# Verify installation
apack version

macOS

# Intel Chip
curl -L https://github.com/model-ci/apack/releases/latest/download/apack-darwin-amd64 -o apack
chmod +x apack
sudo mv apack /usr/local/bin/

# Apple Silicon (M1/M2)
curl -L https://github.com/model-ci/apack/releases/latest/download/apack-darwin-arm64 -o apack
chmod +x apack
sudo mv apack /usr/local/bin/

# Or use Homebrew
brew install model-ci/tap/apack

Windows

# PowerShell download
Invoke-WebRequest -Uri "https://github.com/model-ci/apack/releases/latest/download/apack-windows-amd64.exe" -OutFile "apack.exe"

# Add to PATH environment variable
# Or place directly in system PATH directory

Method 2: Build from Source

# Method 1: Download pre-compiled version from GitHub Releases
wget https://github.com/model-ci/apack/releases/latest/download/apack-linux-amd64
chmod +x apack-linux-amd64
sudo mv apack-linux-amd64 /usr/local/bin/apack

# Method 2: Build from source
git clone https://github.com/model-ci/apack.git
cd apack
make build

Documentation

• User Guide - Detailed usage instructions and best practices
• API Documentation - Complete API interface specifications
• Architecture Design - System architecture and design philosophy
• FAQ - Frequently Asked Questions
• Changelog - Version update history

Examples and Use Cases

• LLM Deployment: Production deployment of large language models
• Research Experiments: Model version management in AI research
• Industrial Applications: Standardized delivery of enterprise AI applications
• Cloud-Native Integration: Model services in Kubernetes environments

Acknowledgments

Special thanks to the following projects and communities:

• Open Container Initiative (OCI) for standard specifications
• llama.cpp community for the inference engine
• Go Language Community for excellent development tools
• All contributors and users for valuable feedback