Skip to main content

Desktop App

The LlamaFarm Desktop App provides a complete local AI environment with visual project management, dataset uploads, chat interface, and built-in model management — no command line required.

Found a bug or have a feature request?

Downloads


Hardware Requirements

Minimum Requirements

To run the desktop app with small models (1-3B parameters like Qwen 1.7B):

ComponentMac (M1+)WindowsLinux
CPUApple M1 or newerIntel i5 / AMD Ryzen 5 (8th gen+)Intel i5 / AMD Ryzen 5 (8th gen+)
RAM8 GB8 GB8 GB
Storage10 GB free10 GB free10 GB free
OSmacOS 12+ (Monterey)Windows 10/11 (64-bit)Ubuntu 22.04+ (tested)
GPUIntegrated (Metal)OptionalOptional

For larger models (7-8B parameters) and better performance:

ComponentMac (M1+)WindowsLinux
CPUApple M1 Pro/Max or M2+Intel i7 / AMD Ryzen 7Intel i7 / AMD Ryzen 7
RAM16 GB+16 GB+16 GB+
Storage50 GB+ SSD50 GB+ SSD50 GB+ SSD
OSmacOS 13+ (Ventura)Windows 11Ubuntu 22.04+
GPUUnified Memory (Metal)NVIDIA RTX 3060+ (8GB+ VRAM)NVIDIA RTX 3060+ (8GB+ VRAM)

Model Memory Requirements

The default model is Qwen 1.7B GGUF (Q4_K_M quantization), which works well on modest hardware.

ModelParametersRAM RequiredVRAM (GPU)Notes
Qwen 1.7B (default)1.7B4 GB2 GBGreat for testing, fast responses
Qwen 3B3B6 GB4 GBBetter quality, still fast
Llama 3.1 8B8B10 GB6 GBHigh quality, needs more resources
Qwen 8B8B10 GB6 GBHigh quality reasoning
Quantization Matters

GGUF models use quantization (Q4_K_M, Q5_K_M, Q8_0) to reduce memory usage. Q4_K_M offers the best balance of quality and speed for most users.


Platform-Specific Notes

Mac (Apple Silicon)

  • Tested on: M1, M1 Pro, M1 Max, M2, M3
  • Acceleration: Uses Metal for GPU acceleration automatically
  • Memory: Unified memory is shared between CPU and GPU — 16GB+ recommended for 8B models
  • Installation: Unzip and drag to Applications folder

Windows

  • Tested on: Windows 10 (21H2+), Windows 11
  • Acceleration: NVIDIA CUDA (if available), otherwise CPU
  • GPU Support: NVIDIA GPUs with CUDA 11.8+ drivers
  • Installation: Run the .exe installer
Windows Defender

Windows Defender may scan the app on first launch. This is normal and should complete within a minute.

Linux

  • Tested on: Ubuntu 22.04 LTS, Ubuntu 24.04 LTS
  • Format: AppImage (portable, no installation needed)
  • Acceleration: NVIDIA CUDA or Vulkan (if available)
  • Dependencies: FUSE required for AppImage
# Make executable and run
chmod +x LlamaFarm-0.0.19.AppImage
./LlamaFarm-0.0.19.AppImage

# If FUSE is not installed:
sudo apt install fuse libfuse2
Other Distributions

While Ubuntu is our primary test platform, the AppImage should work on most modern Linux distributions with glibc 2.31+. Community reports for Fedora, Arch, and Debian are welcome!


Features

The desktop app includes:

  • Visual Project Management — Create, configure, and switch between projects
  • Dataset Uploads — Drag-and-drop file uploads with real-time processing status
  • Chat Interface — Test your AI with full RAG context
  • Model Management — Download, switch, and configure models
  • Built-in Services — No need to run Docker or manage background processes

Troubleshooting

App won't start

  1. Windows: Allow through Windows Defender/Firewall
  2. Linux: Ensure FUSE is installed, check AppImage is executable

Out of memory

  • Close other applications
  • Use a smaller model (Qwen 1.7B instead of 8B)
  • Use higher quantization (Q4_K_M instead of Q8_0)

Model download fails

  • Check internet connection
  • Ensure sufficient disk space
  • Try downloading again — downloads resume automatically

Need help?


Next Steps