Ollama

Ollama runs open-weight models locally on your machine. No API key, no cost, no data leaves your computer. The best option for privacy and free usage.

Setup

# 1. Install Ollama: https://ollama.com
# macOS/Linux:
curl -fsSL https://ollama.com/install.sh | sh

# Windows: download from https://ollama.com/download

# 2. Pull a model
ollama pull llama3.1:8b        # General purpose (8GB RAM)
ollama pull codellama:7b       # Code-focused (8GB RAM)
ollama pull mistral:7b         # Fast, good quality (8GB RAM)
ollama pull llama3.1:70b       # Best quality (48GB+ RAM)

# 3. Arcana auto-detects Ollama
arcana doctor

# 4. Run
arcana run "hello"

Recommended Models

Model	RAM	Best For
`llama3.1:8b`	8GB	General purpose, balanced
`codellama:7b`	8GB	Code generation and completion
`mistral:7b`	8GB	Fast, good quality
`llama3.1:70b`	48GB	Best quality, complex reasoning
`deepseek-coder:6.7b`	8GB	Coding, very capable
`qwen2.5-coder:7b`	8GB	Coding, multilingual

Config File

{
  "provider": "ollama",
  "model": "llama3.1:8b"
}

Environment Variables

Variable	Description
`OLLAMA_HOST`	Ollama server URL (default: `http://localhost:11434`)

No API key required — Ollama runs locally.

Requirements

Model Size	RAM Required	Recommended For
7-8B	8GB+	Everyday coding, quick tasks
13B	16GB+	Better quality, moderate speed
34B	32GB+	High quality, slower
70B	48GB+	Best quality, slowest

Remote Ollama

Run Ollama on a remote machine and connect Arcana:

# On the remote machine:
OLLAMA_HOST=0.0.0.0:11434 ollama serve

# On your local machine:
export OLLAMA_HOST=http://192.168.1.100:11434
arcana run "hello"

Tips

Start with llama3.1:8b — it's the best balance of quality and speed
Use codellama or deepseek-coder for code-focused tasks
Ollama is the only truly free and private option — no data leaves your machine
For better performance, use a GPU (NVIDIA CUDA or Apple Metal)
Run ollama list to see installed models

Last updated: Jul 24, 2026