Your AI runs on a box
you own.
A sovereign assistant on a mini PC on your shelf — not a cloud you rent. One-time hardware, no per-token bill, and your knowledge never leaves the building.
Sovereignty you can point at
The box is the product made physical. Point at the shelf — that is where your data lives, where the AI thinks, and where your compliance posture is grounded.
Your knowledge never leaves the building
The model runs on the appliance, on your network. No prompts, no documents, no embeddings travel to a third-party cloud.
No per-token cloud bill
Inference runs on hardware you bought once. The recurring cost is a thin governance-and-updates tail, not a meter that ticks on every question.
Compliance, made concrete
On-premises by design — so your DSGVO and EU AI Act posture rests on hardware you can point to, with the regulatory layer carried by Cloud Shield rather than a policy you hope holds.
You own it — fork everything but the liability
The open core runs on commodity hardware you control. Nothing locks you to us except the part you would never want to own: the certification and liability.
What turns a mini PC into an appliance
A commodity box becomes a sealed, sovereign AI core because of the operating system it runs. WegentyOS is the hardened, Linux-based OS image — shipped as an installable ISO and flashed onto the machine. Appliance × WegentyOS = AI core.
An OS, not a desktop
A Linux base purpose-built to run open-source models on-device — inference runtime, the connector daemon, the governed answer pipeline and the admin console all present and configured, nothing else. No general-purpose desktop, no telemetry by default.
Sealed by default
No inbound ports. The core is reachable only over the outbound connector tunnel it dials out itself — a minimal attack surface, with hardened defaults rather than a box you have to lock down.
Signed, immutable updates
Image-based, atomic, rollback-capable updates delivered under the maintenance contract — the appliance always converges to a signed, known-good baseline. That is what makes the hardware-and-updates contract a supportable unit instead of a snowflake install.
Open-source like the rest of the stack — everything except Cloud Shield is auditable.
Pick the machine for your knowledge
Every box here can host a governed Wegenty core. Size it to how much you want to run locally — one knowledge base, or several with a 70B model in a single unit.
Transparency — some links here are affiliate links (advertising). Buy through one and we may earn a small commission, at no extra cost to you. We only list machines we would run ourselves. Prices are indicative (June 2026) and move with the market — confirm the live price before you buy.
Entry
7–14B · one knowledge base · from ~€490
NPU/iGPU mini PCs for a single SME with one knowledge base and one public agent.
Minisforum AI X1 Pro
Our pickBest memory headroom in its class — our default box to standardise on, with room to grow.
- Runs
- 7–14B + long context
- Memory
- up to 96 GB (upgradeable)
- Compute
- Ryzen AI 9 HX 370 · Radeon 890M · 50 TOPS
GEEKOM A8
Best valueThe dependable value floor — 3-year warranty and EU fulfilment.
- Runs
- 7–14B
- Memory
- up to 64 GB (upgradeable)
- Compute
- Ryzen 9 8945HS · Radeon 780M
Minisforum UM890 Pro
The cheapest box that qualifies — one SME, one knowledge base.
- Runs
- 7–14B
- Memory
- up to 96 GB (upgradeable)
- Compute
- Ryzen 9 8945HS · Radeon 780M
GMKtec EVO-X1
Highest entry-tier memory bandwidth — more tokens per second.
- Runs
- 7–14B
- Memory
- 32 GB LPDDR5X-7500
- Compute
- Ryzen AI 9 HX 370 · Radeon 890M · 50 TOPS
Beelink SER9 Pro
The most turnkey — ships ready to run a local model out of the box.
- Runs
- 7–14B
- Memory
- 32 GB LPDDR5X-7500
- Compute
- Ryzen AI 9 HX 370 · Radeon 890M · 50 TOPS
ASUS NUC 14 Pro AI
The Intel / vPro pick — the strongest fleet-manageability story.
- Runs
- 7–14B
- Memory
- up to 96 GB
- Compute
- Intel Core Ultra · Arc · 48 TOPS
Performance
30–70B · several knowledge bases · ~€2,400–6,000
Unified-memory boxes (128 GB) or a fast discrete GPU — run a 70B model in one unit.
Minisforum MS-S1 MAX
Our pickRuns a 70B in one box — clustering-ready for several knowledge bases.
- Runs
- 30–70B (clusters higher)
- Memory
- 128 GB unified
- Compute
- Ryzen AI Max+ 395 · Radeon 8060S
RTX 5090 SFF Workstation
FastestFastest for models that fit 32 GB — the CUDA throughput leader of the tier.
- Runs
- ~30B, very fast
- Memory
- 32 GB GDDR7 + system RAM
- Compute
- GeForce RTX 5090 · CUDA · 1,792 GB/s
Beelink GTR9 Pro
Best networking — dual 10 GbE for an on-prem cluster or HA pair.
- Runs
- 30–70B
- Memory
- 128 GB unified
- Compute
- Ryzen AI Max+ 395 · Radeon 8060S
NVIDIA DGX Spark
Full CUDA plus 128 GB — the cloud-consistent path with the strongest software stack.
- Runs
- 30–70B (infer to ~200B; ~405B linked)
- Memory
- 128 GB unified
- Compute
- GB10 Grace Blackwell · full CUDA
Apple Mac Studio M3 Ultra
Most capacityFastest pure inference — silent and low-power. Sold direct (no affiliate link).
- Runs
- 30–100B+
- Memory
- 96–256 GB unified
- Compute
- M3 Ultra · 819 GB/s bandwidth
Framework Desktop
Open referenceThe open, repairable reference box. Direct sales only — buy it yourself, we install.
- Runs
- 30–70B
- Memory
- 128 GB unified
- Compute
- Ryzen AI Max+ 395 'Strix Halo'
Workstations & multi-tenant nodes
For 70B-plus models at speed, or many governed knowledge bases on one box, the appliance steps up to a workstation. These go through a reseller and are quoted to your build — talk to us.
NVIDIA RTX PRO 6000 Workstation
from €12,000Premium 'fast AND it fits' — a 70B at speed with cloud-consistent CUDA.
- Runs
- 70B+ · high concurrency
- Memory
- 96 GB GDDR7 ECC
- Compute
- RTX PRO 6000 Blackwell · CUDA
Dual RTX PRO 6000 Workstation
from €22,000The multi-tenant flagship that still fits a tower — many knowledge bases from one node.
- Runs
- 70–120B · multi-tenant
- Memory
- 192 GB GDDR7 (2× 96 GB)
- Compute
- 2× RTX PRO 6000 Blackwell
NVIDIA DGX Station GB300
On requestThe definitive on-prem multi-tenant appliance — partitions into seven isolated tenants.
- Runs
- 70B → 1T · 7 isolated tenants
- Memory
- 748 GB coherent
- Compute
- GB300 Grace Blackwell Ultra
Prices indicative · June 2026
A consultant supplies it, installs it, keeps it running.
You do not have to source or image anything. A consultant you trust delivers the pre-imaged appliance, installs it in an afternoon, onboards your knowledge, and maintains it — the box, the setup and the support as one relationship.
Are you the consultant?
The appliance is a reseller line on top of the subscription you already run. See how partners earn on hardware.
Earn on hardwareHardware questions
Which box should I pick?
Start from how much you want to run locally. One knowledge base for a single SME → an entry mini PC (the Minisforum AI X1 Pro is our default). Several knowledge bases or a 70B model in one box → a performance unit like the MS-S1 MAX. Many tenants or maximum speed → a workstation. If in doubt, book a demo and we will spec it with you.
Does it really run offline?
Yes. A pure-local core needs no internet at all — the language model and your knowledge base live on the appliance, on your own network, and every answer is formed there. Cloud Shield is the optional public edge: when you want the assistant reachable from outside, the core dials out over the connector tunnel (no inbound ports, no mandatory phone-home), and the same connection carries signed updates. Stay fully local, or add the public edge — that choice is yours.
Which models can it run?
Entry mini PCs comfortably run 7–14B models at 4-bit. The 128 GB unified-memory boxes host a 70B model in a single unit; workstations with 96 GB of VRAM (or more) run 70B-plus at full speed and serve several knowledge bases at once.
Do I have to buy through your links?
No. Buy the machine anywhere you like — the recommendations stand on their own. The affiliate links are simply how we offset the cost of maintaining this guide; using them costs you nothing extra and helps keep the open core free.
Is the hardware part of the subscription?
No — the box is a one-time purchase you own outright. The subscription covers the governance, updates and support around it. That separation is the point: you keep the hardware even if you ever leave.
What else lives on the box?
More than the model. The same core also holds the sovereign, on-core CRM and your customers’ passwordless accounts — the concierge layer that lets the assistant recognise a returning customer and carry their history. Like everything else, those contact records sit on your own core, never a foreign cloud.
Why do the prices change?
The 2026 memory shortage has made AI hardware unusually volatile — some boxes moved by double digits in a single month. We re-verify this list monthly and stamp it with a date, but always confirm the live store price before you order.
Not sure which box fits?
Tell us how much you want to run locally and we will spec the appliance with you — no obligation.