Ollama & Local LLM CLI Master Cheat Sheet

# Verify the local Ollama server is responding
curl -s http://127.0.0.1:11434/api/version

# List running models currently residing in GPU/CPU memory
ollama ps

# Show detailed architectural parameters and system prompt of a model
ollama show --modelfile qwen2.5:7b

Queries daemon health and inspects which neural network weights are currently resident in system VRAM.

# Save the following block into a file named 'Modelfile':
cat << 'EOF' > Modelfile
FROM qwen2.5-coder:7b

# Allocate 16k context window (Requires ~6 GB unified memory / VRAM)
PARAMETER num_ctx 16384

# Low temperature for deterministic, bug-free refactoring
PARAMETER temperature 0.2

# Penalize repetition in long code generation loops
PARAMETER repeat_penalty 1.15
PARAMETER top_p 0.9

# Define strict developer system prompt
SYSTEM """
You are an expert systems engineer. You write concise, production-ready,
and memory-safe code. Always adhere to strict typing, error handling,
and clean code principles. Do not include conversational fluff.
"""
EOF

# Build and register the customized model into Ollama
ollama create code-expert -f ./Modelfile

Generates a specialized local coding model with custom context length, sampling parameters, and system constraints.

# Request guaranteed JSON-formatted output without streaming
curl -s http://127.0.0.1:11434/api/chat -d '{
  "model": "qwen2.5:7b",
  "messages": [
    {
      "role": "system",
      "content": "Extract data into JSON with keys: name, role, language."
    },
    {
      "role": "user",
      "content": "Alice is a DevOps engineer who specializes in Rust."
    }
  ],
  "format": "json",
  "stream": false
}' | jq .message.content

Passes format: "json" to enforce constrained decoding, returning a pristine JSON object.

# Generate high-dimensional vector embeddings for RAG pipelines
curl -s http://127.0.0.1:11434/api/embeddings -d '{
  "model": "nomic-embed-text",
  "prompt": "Local in-browser computing ensures complete privacy."
}' | jq '.embedding | length'

Extracts dense vector embeddings locally without sending sensitive internal knowledge to cloud embedding APIs.

# Linux / macOS: Configure Ollama daemon for multi-user concurrency and warm VRAM retention
export OLLAMA_NUM_PARALLEL=4          # Handle 4 parallel requests concurrently
export OLLAMA_MAX_LOADED_MODELS=2     # Retain up to 2 models resident in VRAM
export OLLAMA_KEEP_ALIVE="24h"        # Keep models loaded in memory for 24 hours
export OLLAMA_FLASH_ATTENTION=1       # Enable Flash Attention for lower VRAM overhead

# Start the Ollama server with customized parameters
ollama serve

Tunes server concurrency, prevents repetitive cold-start loading penalties, and activates Flash Attention.

# Open the systemd override editor for the Ollama service
sudo systemctl edit ollama.service

# Add the following configuration block:
[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_ORIGINS=https://my-internal-dashboard.example.com"
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_KEEP_ALIVE=1h"
Restart=always
RestartSec=3

# Reload systemd and restart the daemon
sudo systemctl daemon-reload
sudo systemctl restart ollama

Applies persistent environment variables and automatic restart watchdogs to the production Ollama daemon.