LIVE TAPE—
★OpenClaw 88,412 stars·⚠CVE-2026-25898 disclosed (HIGH, Hermes)·↻Hermes Agent v2026.4.7 published·★Hermes Agent +182 stars (last hour)·↻OpenClaw v2026.4.6 — credential vault hardening·⚠CVE-2026-26133 patched (NanoClaw)·•Pi 5 16GB rumoured for Q3 — recheck guidance·★Nanobot +47 stars (last hour)·↻ZeroClaw v0.4.2 — Apple container fixes·•Mac Mini M4 wins quarterly hardware survey·★OpenClaw 88,412 stars·⚠CVE-2026-25898 disclosed (HIGH, Hermes)·↻Hermes Agent v2026.4.7 published·★Hermes Agent +182 stars (last hour)·↻OpenClaw v2026.4.6 — credential vault hardening·⚠CVE-2026-26133 patched (NanoClaw)·•Pi 5 16GB rumoured for Q3 — recheck guidance·★Nanobot +47 stars (last hour)·↻ZeroClaw v0.4.2 — Apple container fixes·•Mac Mini M4 wins quarterly hardware survey·
PocketClawvol. 1 · 2026
← all comparisons
comparison

Raspberry Pi 5 vs Intel NUC 13

Sub-€150 ARM SBC versus mid-tier x86 mini PC. The crossover point is around 7B model size.

Updated April 2026 · Same VPS, same five tasks, same scoring rubric.

Side-by-side

AxisRaspberry Pi 5Intel NUC 13
Setup timeRaspberry Pi OS Bookworm flash, Docker, Hermes Agent — 60-90 minutes from box to running agent. Bring your own SD card or NVMe HAT.Ubuntu 24.04 LTS install, Docker, Hermes Agent — 45 minutes from box to running agent. Internal NVMe ships with the unit.
Security modelARM Cortex-A76 has Pointer Authentication; useful but not a panacea. No hardware TPM by default.Intel TXT and TPM 2.0 supported. Intel Boot Guard for verified boot. Hardware-rooted attestation possible.
Model supportComfortable up to 3B Q4 (~7 tok/s on qwen2.5-coder-3B). 7B starts hurting at 4 tok/s. 14B+ unrealistic.Comfortable up to 7B Q4 (~11 tok/s on qwen2.5-coder-7B). 14B Q4 runs at 6 tok/s — slow but usable. 32B is unrealistic without GPU.
Cost€138 hardware. €4.30/month total cost of ownership including electricity.€620 hardware. €25/month total cost of ownership including electricity.
EcosystemRaspberry Pi has the largest single-board-computer community on Earth. Every problem has been solved by someone, somewhere.Standard x86. Anything that runs on a Linux laptop runs on a NUC.
Best forEdge classification, agent orchestration, small-model workloads. Or anyone learning self-hosted AI cheaply.Mid-volume self-hosted AI on 7B models, or workloads where you'd otherwise want a laptop minus the screen.

Verdict

Pi 5 if your workload fits inside 3B. NUC 13 if you need 7B with margin. Both lose to Mac Mini M4 on tokens-per-watt; both win against the GPU box on idle power and price. Choose by model size, not headline brand.

Notes

  • Pi 5 needs a fan — passive cooling alone throttles in five minutes of LLM work.
  • NUC 13's Iris Xe iGPU helps a little on inference; Vulkan or SYCL backend in llama.cpp can pull 20% more tokens/s.
  • If you're going to add NVMe, a Mac Mini M4 24 GB is often a better dollar-for-dollar buy than a NUC + storage.

Going deeper

For the full landscape report including hosting economics, security posture and regulatory context, see the 2026 landscape report. For the OpenClaw-specific history, see the complete OpenClaw timeline.

New comparison requests are welcome — subscribe and reply to any edition with your short-list.