Skip to main content

​
Overview

The cactus run command opens an interactive playground for any supported model. If the model isn’t already downloaded, Cactus will automatically fetch it from HuggingFace.

​
Syntax

cactus run <model> [flags]

​
Arguments

  • <model> - Model name (e.g., qwen-2.5-1.5b, llama-3.2-1b, phi-4)

​
Flags

​
—precision

Set the quantization precision level: 
cactus run <model> --precision INT4|INT8|FP16
Default: INT4 Options:
  • INT4 - 4-bit quantization (smallest size, fastest)
  • INT8 - 8-bit quantization (balanced)
  • FP16 - 16-bit floating point (highest quality)

​
—token

Provide a HuggingFace API token for gated models: 
cactus run <model> --token <your-hf-token>
Required for models like Llama that require access approval.

​
—reconvert

Force reconversion of the model from source weights: 
cactus run <model> --reconvert
Useful when model format has been updated or conversion failed previously.

​
Examples

# Run Qwen with default settings (INT4)
cactus run qwen-2.5-1.5b

​
Interactive Playground

Once the model loads, you’ll enter an interactive chat interface: 
┌─────────────────────────────────────────────┐
│ Cactus Playground - qwen-2.5-1.5b          │
│ Precision: INT4                             │
└─────────────────────────────────────────────┘

You: Hello! What can you help me with?

Assistant: I'm an AI assistant running locally on
your device. I can help with coding, writing,
analysis, and general questions...

​
Model Auto-Download

If the model isn’t cached locally, cactus run will:
  1. Download the model from HuggingFace
  2. Convert it to Cactus format with the specified precision
  3. Cache it in ./weights for future use
  4. Launch the interactive playground

​
See Also

Download Command

Pre-download models without running them

Model Library

Browse all supported models