🌾 LLaMA Farm CLI

Deploy AI models, agents, and databases into single deployable binaries - no cloud required.

Installation

1npm install -g @llamafarm/llamafarm

Quick Start

1# Deploy a model
2llamafarm plant llama3-8b
3
4# Deploy with optimization
5llamafarm plant llama3-8b --optimize
6
7# Deploy to specific target
8llamafarm plant mistral-7b --target raspberry-pi
9
10# Development/Testing (no model download)
11llamafarm plant llama3-8b --mock

Complete Workflow Example

1# 1. Plant - Configure your AI deployment
2llamafarm plant llama3-8b \
3  --device mac-arm \
4  --agent chat-assistant \
5  --rag \
6  --database vector
7
8# 2. Bale - Compile to single binary
9llamafarm bale ./.llamafarm/llama3-8b \
10  --device mac-arm \
11  --optimize
12
13# 3. Harvest - Deploy anywhere
14llamafarm harvest llama3-8b-mac-arm-v1.0.0.bin --run
15
16# Or just copy and run directly (no dependencies needed!)
17./llama3-8b-mac-arm-v1.0.0.bin

Features

• 🎯 One-Line Deployment - Deploy complex AI models with a single command
• 📦 Zero Dependencies - Compiled binaries run anywhere
• 🔒 100% Private - Your data never leaves your device
• ⚡ Lightning Fast - 10x faster than traditional deployments
• 💾 90% Smaller - Optimized models use fraction of original size

Commands

plant

Deploy a model to create a standalone binary.

1llamafarm plant <model> [options]
2
3Options:
4  --target <platform>    Target platform (mac, linux, windows, raspberry-pi)
5  --optimize            Enable size optimization
6  --agent <name>        Include an agent
7  --rag                 Enable RAG pipeline
8  --database <type>     Include database (vector, sqlite)

Examples

1# Basic deployment
2llamafarm plant llama3-8b
3
4# Deploy with RAG and vector database
5llamafarm plant mixtral-8x7b --rag --database vector
6
7# Deploy optimized for Raspberry Pi
8llamafarm plant llama3-8b --target raspberry-pi --optimize
9
10# Deploy with custom agent
11llamafarm plant llama3-8b --agent customer-service

bale

🎯 The Baler - Compile your deployment into a single executable binary.

1llamafarm bale <project-dir> [options]
2
3Options:
4  --device <platform>   Target platform (mac, linux, windows, raspberry-pi)
5  --output <path>       Output binary path
6  --optimize <level>    Optimization level (none, standard, max)
7  --sign               Sign the binary for distribution
8  --compress           Extra compression (slower but smaller)

The Baler packages everything into a single binary:

• 🧠 Quantized model (GGUF format)
• 🤖 Agent configuration & code
• 🗄️ Embedded vector database
• 🌐 Web UI
• 🚀 Node.js runtime
• 🔧 Platform-specific optimizations

Supported Platforms:

• mac / mac-arm / mac-intel - macOS with Metal acceleration
• linux / linux-arm - Linux with CUDA support
• windows - Windows with DirectML/CUDA
• raspberry-pi - Optimized for ARM devices
• jetson - NVIDIA Jetson edge devices

Typical Binary Sizes:

• 7B models: 4-8GB (depending on quantization)
• 13B models: 8-13GB
• Mixtral: 25-45GB

Bale Examples

1# Standard compilation
2llamafarm bale ./.llamafarm/llama3-8b --device mac-arm
3
4# Optimized for size
5llamafarm bale ./.llamafarm/llama3-8b --device raspberry-pi --optimize max --compress
6
7# Enterprise deployment with signing
8llamafarm bale ./.llamafarm/mixtral --device linux --sign --output production.bin

harvest

Deploy and run a compiled binary.

1llamafarm harvest <binary-or-url> [options]
2
3Options:
4  --run                 Run immediately after deployment
5  --daemon             Run as background service
6  --port <number>      Override default port
7  --verify             Verify binary integrity

Configuration

Create a llamafarm.yaml file for advanced configurations:

1name: my-assistant
2base_model: llama3-8b
3plugins:
4  - vector_search
5  - voice_recognition
6data:
7  - path: ./company-docs
8    type: knowledge
9optimization:
10  quantization: int8
11  target_size: 2GB

Then build:

1llamafarm build

Requirements

• Node.js 18+
• 8GB RAM (minimum)
• 10GB free disk space

Documentation

For full documentation, visit https://docs.llamafarm.ai

Support

• 📖 Documentation
• 💬 Discord Community
• 🐛 Issue Tracker

Baler FAQ

Q: Can I run the binary on a different OS than where I compiled it? A: No, you need to compile for each target platform. Use --device to specify the target.

Q: How much disk space do I need? A: During compilation, you need ~3x the final binary size. The final binary is typically 4-8GB for 7B models.

Q: Can I update the model without recompiling? A: No, the model is embedded in the binary. This ensures zero dependencies but means updates require recompilation.

Q: Does the binary need internet access? A: No! Everything runs completely offline once deployed.

License

MIT © LLaMA Farm Team