Spec sheets say "67 TOPS". Marketing pages say "runs AI at the power of a lightbulb". Neither tells you what you actually get when you type a prompt. So we took a production ClawBox off the shelf — the same NVIDIA Jetson Orin Nano 8GB unit we ship — and measured it.
No lab conditions, no cherry-picking: Ollama, warm runs, real wattage sampled from the board itself. Here are the numbers, the methodology, and an honest read of what they mean for running an always-on assistant at home.
The results
| Model | Tokens/sec | Model load time |
|---|---|---|
| llama3.2:1b | 44.9 | 0.4s |
| gemma2:2b | 25.6 | 0.5s |
| qwen2.5:3b | 22.6 | 0.4s |
| phi3:mini (3.8B) | 22.0 | 0.1s |
Power and thermals during the run (tegrastats, 176 samples):
- Peak total board power: 19.2 W — average ~11.4 W
- Compute rail (VDD_CPU_GPU_CV) peak: 9.0 W
- Peak junction temperature: 61.8°C — no thermal throttling
- RAM in use during inference: ~4.8 of 7.6 GB
- GPU utilization (GR3D): 99% during generation
The one-line takeaway: a ClawBox runs real 1–4B-class local models at 22–45 tokens/sec while drawing under 20 watts, peak. For context, that average draw is less than most LED desk lamps, and roughly 1/40th of a gaming PC under load.
Methodology (so you can reproduce it)
- Device: production ClawBox unit — NVIDIA Jetson Orin Nano 8GB, JetPack, aarch64, NVMe storage. Nothing overclocked, stock power mode.
- Runner: Ollama, the same way most self-hosters run local models.
- Procedure: each model pulled, loaded, given a warm-up generation, then measured on a warm run with
num_predict=200. Tokens/sec is Ollama's owneval_ratefor the generation phase. - Power/thermals:
tegrastatssampled throughout the entire benchmark session (176 samples), capturing total board power, per-rail power, and junction temperature.
If you own any Jetson Orin Nano (or a ClawBox), you can replicate this in ten minutes. If you get meaningfully different numbers, we'd genuinely like to hear about it.
What 22–45 tokens/sec actually feels like
Reading speed for most people is 5–8 tokens/sec. Everything in this table generates faster than you can read it:
- llama3.2:1b at ~45 tok/s — instant-feeling. This is the class of model you use for routing, summarizing notifications, tagging email, quick rewrites. On-device, private, effectively free to run all day.
- 2–3B models at ~22–26 tok/s — comfortable chat speed. Good enough for drafting, local Q&A over your files, and the private "first pass" work an assistant does constantly in the background.
- phi3:mini (3.8B) at 22 tok/s — the practical ceiling of what fits comfortably in 8GB alongside the rest of the stack. Solid reasoning for its size, still faster than reading speed.
The honest part: this is not a machine for running 70B models locally. Nothing at this power class is. The pattern that actually works — and the one OpenClaw is built around — is hybrid: small local models handle the private, high-frequency, always-on work; cloud models (Claude, GPT) handle the heavy reasoning when you ask for it. The benchmark shows the local half of that split is genuinely usable, not a gimmick.
Why the wattage number matters more than the speed number
An always-on assistant is only practical if you can leave it on. At ~11.4W average, a ClawBox running 24/7 uses about 100 kWh per year — roughly €25–35 of electricity in most of Europe, a rounding error next to a desktop PC doing the same job. That's the difference between an assistant you switch on when you need it and one that's just there — watching your inbox, running scheduled jobs, answering on Telegram at 3 AM.
And at 62°C peak with no throttling, it does this silently on a shelf, not on a machine you hear across the room.
Where this fits in the hardware decision
If you're weighing your options for running OpenClaw, these numbers slot into the bigger picture:
- The full spec breakdown — minimum vs recommended, what actually consumes RAM — is on our OpenClaw hardware requirements page.
- The honest comparison of your real choices — mini PC, DIY Jetson dev kit, Mac mini, ClawBox — is in Best Hardware to Run OpenClaw in 2026.
- If you want the machine these numbers came from, pre-configured with OpenClaw and ready in about 5 minutes: ClawBox, €549.
We'll keep publishing measured numbers as JetPack and Ollama releases move — this page reflects a production unit as shipped in July 2026.
Benchmark run 2026-07-21 on a production ClawBox. Raw data: internal record; methodology above is complete enough to reproduce independently.