The 13-Watt Miracle: Mac Mini M4
We just ran the numbers, and they don't look real.
At The Neural Lab, we are used to seeing power meters spin like centrifuges. Our RTX 5070 test bench pulls 220 Watts under full load. Our RTX 4060 Laptop ("The People's Champion") sips a modest 90-110 Watts when generating tokens.
Then we plugged in the Mac Mini M4.
The Benchmark
We ran Llama-3.1-8B (Q4_K_M) through our standard thermal soak test.
- Model: Llama 3.1 8B
- Context: 4096 tokens
- Duration: 30 minutes continuous inference
The result? 21.39 tokens/second at 13.15 Watts (GPU Power).
Context Matters: GPU vs. System Power
To be technically precise: that 13.15W figure is the power draw of the GPU slice of the M4 So, measured via Apple's powermetrics tool. It does not include the CPU cores, the DRAM, or the rest of the package.
However, even if you double it to account for system overhead, the efficiency is staggering.
- NVIDIA RTX 4060 Laptop: ~100W for ~50 t/s (~2 Watts per token/s).
- M4 Mac Mini: ~13W for ~21 t/s (~0.6 Watts per token/s).
The M4 is roughly 3x more efficient per token generated than the current standard for mobile inference.
The "Always-On" Advantage
Why does this matter? Because of Idle Cost.
If you are building a home intelligence server—a machine that listens for Home Assistant voice commands, summarizes your emails at 6 AM, and categorizes your photos—you can't have a rig that idles at 100W and spikes to 300W. That heat (and electricity bill) adds up.
The Mac Mini M4 runs cool, practically silent, and consumes less power than a lightbulb while running a model that is smarter than most humans.
The Verdict
The Mac Mini M4 isn't a race car. It will lose a drag race against an RTX 5060 every single time (especially in ingestion speed). But it's not a race car. It's a Prius Prime. It's the perfect daily driver for the "Always-On" era of local AI.
- Prompt Evaluation: 231 t/s (Slow)
- Generation: 21 t/s (Usable)
- Power: 13W (Miraculous)
If you want to build a cluster of agents that run 24/7, this is your hardware.