Michael Limberger

Performance Tuning And End-To-End Testing

Performance tuning and end-to-end testing

Everything is installed, configured, and hardened. This last section covers getting more out of Apple Silicon hardware, then verifying the full stack before you rely on it.

Keeping the model warm

By default, Ollama unloads models from memory after 5 minutes of inactivity. The next request triggers a cold start that takes 10 to 30 seconds while the model reloads into GPU memory. For an always-ready assistant, keep the model loaded:

export OLLAMA_KEEP_ALIVE="24h"

That tells Ollama to keep models loaded for 24 hours after the last request. Put it in your shell profile (~/.zshrc) to make it permanent.

You can also pre-warm the model after a reboot so it is ready before your first question:

Show me

curl http://localhost:11434/api/generate \
          -d '{"model":"qwen2.5-coder:14b","prompt":"warmup","keep_alive":"24h"}'

Flash Attention and KV cache

Two environment variables that improve memory efficiency:

export OLLAMA_FLASH_ATTENTION=1
        export OLLAMA_KV_CACHE_TYPE="q8_0"

Flash Attention reduces attention memory overhead. The quantized KV cache (q8_0) saves roughly half the memory used for conversation context. That matters most on 16GB machines where every megabyte counts.

Add both to your shell profile.

Custom Modelfile for locked settings

Instead of relying on defaults, create a Modelfile that locks your preferred context window.

Create a file called Modelfile with these contents:

FROM qwen2.5-coder:14b
        PARAMETER num_ctx 16384
        PARAMETER num_gpu 99

Then build a custom model from it:

Show me

ollama create openclaw-coder -f Modelfile

Reference openclaw-coder in your OpenClaw config instead of qwen2.5-coder:14b. That keeps the context window at 16K regardless of Ollama's defaults.

On 16GB RAM, use 8192 instead of 16384 for num_ctx. On 32GB+, you can push to 32768.

num_gpu 99 means "use all GPU layers." That is already the default on Apple Silicon, but making it explicit does not hurt.

Context window and OpenClaw's overhead

Something that is easy to miss: OpenClaw's system prompt alone consumes about 17,000 tokens. The agent also uses compaction and memory files for long conversations, but tool-heavy workflows fill context fast.

On a 16GB machine with num_ctx set to 8192, raw arithmetic looks grim: the system prompt alone can exceed 8K.

In practice, OpenClaw compacts the system prompt and manages context more carefully than raw token counting suggests. Still, keep your context window as large as your RAM allows.

On 16GB, use 8192 and accept that complex multi-step tasks may hit the ceiling. On 24GB+, use 16384 to 32768.

Monitoring during operation

Keep an eye on these while using OpenClaw:

Show me

GPU utilization:

          ollama ps

          Should show "100% GPU" in the processor column. If it shows any
          CPU split, the model is partially offloaded and will be slower.

        Memory pressure:

          Open Activity Monitor, click Memory tab. The "Memory Pressure"
          graph at the bottom should stay green. Yellow means you're starting
          to swap. Red means pain.

        Token speed:

          ollama run qwen2.5-coder:14b --verbose "test"

          Look for "eval rate: XX.XX tokens/s" in the output. This is your
          generation speed. Compare against the expected numbers for your
          chip (8-12 for M1, 15-22 for Pro, 25-40 for Max).

Thermal throttling on laptops

M-series chips sustain peak GPU clocks for about 90 seconds before thermal throttling kicks in. On a MacBook, sustained inference can run maybe 10 to 15% slower after the first couple of minutes.

For a demo or extended use:

Elevate the laptop for better airflow under the chassis.
        Use a desk fan pointed at the base.
        Avoid clamshell mode (closed lid with external display) unless
        you have good ventilation around the hinge area.

This is minor for most uses. It mainly matters during long conversations where the model runs continuously for minutes at a time.

End-to-end verification checklist

Run these in sequence to confirm everything works:

Show me

Step 1: Ollama is running and model is loaded

          curl http://localhost:11434/
          ollama list | grep qwen2.5-coder
          ollama ps

        Step 2: OpenClaw gateway is up

          openclaw gateway status
          curl -s http://localhost:18789/health

        Step 3: OpenClaw can see the Ollama model

          openclaw models list
          openclaw models status

        Step 4: Full diagnostic

          openclaw doctor --fix

        Step 5: Security audit (do not skip this)

          openclaw security audit --deep

        Step 6: Verify no cloud fallback

          openclaw logs --follow

          In another terminal, send a test message through TUI or Dashboard.
          Watch logs for any "anthropic" or "openai" references.

Test conversations

Open the Dashboard or TUI and try these in order:

Test 1: "What day is it today?"
        Confirms the model responds at all. If nothing comes back, check
        the troubleshooting section.

        Test 2: "List the files in my current directory"
        Confirms tool calling works. The agent should actually execute a
        shell command and return real filesystem contents. Not describe
        what it would do. Actually do it.

        Test 3: "Create a file called test.txt with 'Hello from OpenClaw'
        inside it, then read it back to me"
        Confirms multi-step tool use and file I/O.

        Test 4: "What's in my ~/.openclaw/openclaw.json file?"
        Confirms file reading works and the agent has appropriate
        filesystem access.

If all four pass, your stack is working correctly.

Group session tips

If you are setting this up for multiple people at once (meetup, workshop, team session), prepare in advance:

Pre-download models on every machine. Thirty people pulling 9GB
        simultaneously over the same Wi-Fi will saturate any connection. If
        your venue allows it, distribute the model file via USB drives.

        Set OPENCLAW_DISABLE_BONJOUR=1 on every machine. Without this,
        every OpenClaw instance broadcasts its presence on the network. With
        thirty instances on the same Wi-Fi, you get mDNS collisions, device
        discovery confusion, and potentially crashes.

        Use the Web Dashboard, not messaging platforms. For a group demo, the
        Dashboard is simpler and more reliable. No external service
        dependencies, no pairing codes, no account issues.

        Budget timing realistically:

          Ollama install:                    2 minutes
          Model pull (pre-downloaded):       0 minutes
          Model pull (live, 100 Mbps):       12-15 minutes
          OpenClaw install:                  1-3 minutes
          Configuration + onboarding:        5 minutes
          Security audit + hardening:        3 minutes
          First successful conversation:     under 1 minute

        Pre-prepared attendees: about 15 minutes total.
        Downloading models live: 30+ minutes. Plan accordingly.

+----------------------------------------------------------+
|  The two things that matter most:                        |
|                                                          |
|  1. Pre-download the model. Nothing kills momentum like  |
|     thirty people waiting for a 9GB download.            |
|                                                          |
|  2. Run the security audit. Not later. Not eventually.   |
|     Before you use it. Before you demo it.               |
+----------------------------------------------------------+

That is the complete guide. You have a fully local AI assistant on your Mac's GPU, configured to avoid cloud contact, hardened against the known attack paths, and tuned for your hardware.

Whether you use it daily or just wanted to understand how local AI works, you now have the pieces. Use it wisely. And if the lobsters start recruiting you into Crustafarianism, maybe step away from the terminal for a bit.

Stay curious, be careful, and have fun.