Pi 5 vs x86 mini-PC for self-hosted AI: a real-numbers comparison

We get this question every week. The short answer is “both work, here’s how to pick.” The long answer is below, with numbers.

What we’re measuring

A self-hosted AI assistant box has three workloads:

1. The agent + channels — Node.js (OpenClaw) or Python (Hermes/Khoj) running 24/7. Memory matters more than CPU.
2. Local LLM inference (optional) — llama.cpp running a baked GGUF on the air-gapped tier, or as fallback. CPU matters a lot.
3. Indexing (Khoj only) — embedding documents into a vector store. CPU + RAM both matter.

Different hardware tilts different workloads.

The contenders

Pi 5 8 GB (~$80)

• Cortex-A76 quad-core @ 2.4 GHz, NEON-capable.
• LPDDR4X-4267, 8 GB.
• microSD or NVMe via the M.2 hat.
• 5 V 5 A PSU required.
• Idle: ~5 W. Load: ~9 W.

Pi 5 16 GB (~$120)

• Same SoC as 8 GB, double RAM.
• Same power profile.
• 16 GB headroom is what unlocks LFM2-2.6B air-gapped and Khoj indexing without thrash.

Beelink-class mini-PC (~$200, varies)

• Intel N100 or Ryzen-class APU (varies by SKU).
• 16 GB DDR4 or DDR5, NVMe storage.
• ~10 W idle, ~25 W under load.
• AVX2 (and sometimes AVX-512) for llama.cpp speedups.

(There’s also Apple silicon Mac Mini — overkill for the agent role, but if you have one already, it’s excellent.)

Numbers we actually measured

These are rough numbers from our test rigs, not lab-grade benchmarks. Your mileage varies with thermal solution, OS load, SD vs NVMe storage. They’re directionally accurate.

Agent boot + idle

Workload	Pi 5 8 GB	Pi 5 16 GB	Beelink N100
Boot to deputyos.local responsive	~70 s	~70 s	~30 s
OpenClaw memory footprint	~280 MB	~280 MB	~280 MB
Hermes memory footprint	~420 MB	~420 MB	~420 MB
Khoj memory + idle index	~700 MB	~700 MB	~600 MB
Idle CPU	<1%	<1%	<1%

For agent-only workloads, the boot time is the only meaningful Pi-vs-mini-PC difference.

Local LLM inference (llama.cpp, Q4_K_M)

Model	Pi 5 8 GB	Pi 5 16 GB	Beelink N100
LFM2-350M (lean)	~22 t/s	~22 t/s	~70 t/s
LFM2-1.2B (standard)	~9 t/s	~9 t/s	~24 t/s
LFM2-2.6B (rich)	refuses to load (OOM)	~5 t/s	~12 t/s
Qwen2.5-Coder-1.5B	~6 t/s	~6 t/s	~15 t/s
Llama-3.2-1B (legacy)	~10 t/s	~10 t/s	~28 t/s

A Pi 5 8 GB literally cannot fit LFM2-2.6B in its KV-cache budget alongside the agent — deputyOS refuses to bake that tuple. Pi 5 16 GB can; it’s slow but works for chat. The Beelink is in a different league because of AVX2 and DDR5 bandwidth.

Khoj indexing (1000 markdown notes, ~5 MB total)

Pi 5 8 GB	Pi 5 16 GB	Beelink N100
~4 min	~3 min	~45 sec

Re-indexing a personal-knowledge base on a Pi is a coffee break. On a mini-PC it’s a stretch. If your Khoj corpus is large (~10k+ docs), the mini-PC is meaningfully nicer.

Long-running power draw

Hardware	Idle (W)	Typical chat (W)	Peak (W)	$/year @ $0.15/kWh, 24/7 idle
Pi 5 8 GB	~5	~6	~9	~$7
Pi 5 16 GB	~5	~6	~9	~$7
Beelink N100	~10	~12	~25	~$14

In US markets, the annual electricity difference is one cheap lunch. In high-cost markets (UK / EU) it’s a slightly more meaningful but still small number.

When Pi 5 is the right pick

You’re running OpenClaw or Hermes against a cloud LLM (you’ll pay for tokens, not for inference). You want the box to be silent and small. You’re starting out and want to keep the entry cost low.

You can always add a Pi 5 16 GB later (or a mini-PC) if you outgrow the 8 GB; the deputyOS image transfer is straightforward and your conversation history backs up cleanly.

When mini-PC is the right pick

You want air-gapped operation with the rich tier (LFM2-2.6B + Qwen-Coder). You run Khoj heavily and re-index often. You’re building a multi-profile setup (OpenClaw on one box, Khoj on another) and want both to be the same hardware class.

The Beelink-class N100 is, dollar-for-dollar, the best $200 you can spend on always-on self-hosted compute right now.

When both is the right pick

This is what most “serious self-hosters” end up doing:

• Pi 5 (8 GB) → OpenClaw + channels. Always-on. Front door for messages.
• Beelink → Khoj + air-gapped Hermes. Heavier workloads. Indexes your files.

The two devices share a LAN; channels route to the Pi; complex queries route to the Beelink. deputyOS doesn’t have a built-in “federate two devices” feature yet (on the roadmap), but you can do it manually via a skill that delegates to the second device’s /chat endpoint.

Storage matters

SD cards on a Pi 5 work but are the slow part. Two upgrades:

• NVMe via the M.2 hat (~$25 for the hat + $20 for a 256 GB NVMe). Boot in 30 seconds, agent feels instantly responsive.
• USB3 SSD (~$30 for a 256 GB drive + USB3 enclosure). Slightly less convenient than NVMe; comparable speed.

Either upgrade is the single biggest perceived-speed change you can make to a Pi 5.

What hardware we recommend right now

• First-time user, just want to chat: Pi 5 8 GB + SD card. Total: ~$95.
• Daily-driver assistant + a few skills: Pi 5 8 GB + NVMe + M.2 hat. Total: ~$130.
• Power user, air-gapped enthusiast: Pi 5 16 GB + NVMe. Total: ~$170.
• Khoj-heavy / inference enthusiast: Beelink-class N100 16 GB. Total: ~$200.
• The combo: Pi 5 + Beelink. Total: ~$300.

What about Mac Mini / NUC / X3D desktops?

All work fine and the agent will be very responsive. Overkill for the agent role — you’ll see <0.1% CPU on a recent Mac Mini. If you already have one running 24/7, definitely use it. The desktop launcher path is the cleanest way to run deputyOS on a Mac without dedicating the hardware.

The picker has images for Pi 5, Pi 4, x86 mini-PC, generic arm64, and every cloud option. Once you’ve picked hardware, the 10-minute Pi 5 walkthrough applies equally to the mini-PC (just pick x86_64-mini-pc in the picker).