• Mac mini with Apple Silicon (M1/M2/M3/M4/M5)
• At least 24GB unified memory for Gemma 4 26B
• macOS with Homebrew installed

Install the Ollama macOS app via Homebrew cask (includes auto-updates and MLX backend):

brew install --cask ollama-app

This installs:

• Ollama.app in /Applications/
• ollama CLI at /opt/homebrew/bin/ollama

The Ollama icon will appear in the menu bar. Wait a few seconds for the server to initialize.

Verify it's running:

This downloads ~17GB. Verify:

ollama list
# NAME          ID              SIZE     MODIFIED
# gemma4:26b    5571076f3d70    17 GB    ...

ollama run gemma4:26b "Hello, what model are you?"

Check that it's using GPU acceleration:

ollama ps
# Should show CPU/GPU split, e.g. 14%/86% CPU/GPU

Click the Ollama icon in the menu bar > Launch at Login (enable it).

Alternatively, go to System Settings > General > Login Items and add Ollama.

Create a launch agent that loads the model into memory after Ollama starts and keeps it warm:

cat << 'EOF' > ~/Library/LaunchAgents/com.ollama.preload-gemma4.plist
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.ollama.preload-gemma4</string>
    <key>ProgramArguments</key>
    <array>
        <string>/opt/homebrew/bin/ollama</string>
        <string>run</string>
        <string>gemma4:26b</string>
        <string></string>
    </array>
    <key>RunAtLoad</key>
    <true/>
    <key>StartInterval</key>
    <integer>300</integer>
    <key>StandardOutPath</key>
    <string>/tmp/ollama-preload.log</string>
    <key>StandardErrorPath</key>
    <string>/tmp/ollama-preload.log</string>
</dict>
</plist>
EOF

Load the agent:

launchctl load ~/Library/LaunchAgents/com.ollama.preload-gemma4.plist

This sends an empty prompt to ollama run every 5 minutes, keeping the model warm in memory.

By default, Ollama unloads models after 5 minutes of inactivity. To keep them loaded forever:

launchctl setenv OLLAMA_KEEP_ALIVE "-1"

Then restart Ollama for the change to take effect.

Note: This environment variable is session-scoped. To persist across reboots, add export OLLAMA_KEEP_ALIVE="-1" to your ~/.zshrc, or set it via a dedicated launch agent.

# Check Ollama server is running
ollama list

# Check model is loaded in memory
ollama ps

# Check launch agent is registered
launchctl list | grep ollama

Expected output from ollama ps:

NAME          ID              SIZE     PROCESSOR          CONTEXT    UNTIL
gemma4:26b    5571076f3d70    20 GB    14%/86% CPU/GPU    4096       Forever

Ollama exposes a local API at http://localhost:11434. Use it with coding agents:

# Chat completion (OpenAI-compatible)
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma4:26b",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

# Remove the preload agent
launchctl unload ~/Library/LaunchAgents/com.ollama.preload-gemma4.plist
rm ~/Library/LaunchAgents/com.ollama.preload-gemma4.plist

# Uninstall Ollama
brew uninstall --cask ollama-app

On Apple Silicon, Ollama automatically uses Apple's MLX framework for faster inference — no manual configuration needed. M5/M5 Pro/M5 Max chips get additional acceleration via GPU Neural Accelerators. M4 and earlier still benefit from general MLX speedups.

Ollama now leverages NVIDIA's NVFP4 format to maintain model accuracy while reducing memory bandwidth and storage requirements for inference workloads. As more inference providers scale inference using NVFP4 format, this allows Ollama users to share the same results as they would in a production environment. It further opens up Ollama to run models optimized by NVIDIA's model optimizer.

• Lower memory utilization: Ollama reuses its cache across conversations, meaning less memory utilization and more cache hits when branching with a shared system prompt — especially useful with tools like Claude Code.
• Intelligent checkpoints: Ollama stores snapshots of its cache at intelligent locations in the prompt, resulting in less prompt processing and faster responses.
• Smarter eviction: Shared prefixes survive longer even when older branches are dropped.

• Memory: Gemma 4 26B uses ~20GB when loaded. On a 24GB Mac mini, this leaves ~4GB for the system — close memory-heavy apps before running.