Your AI.
On your ground.

Your hardware. Your data. Unlimited usage.

Compare systems
Private on your network
Complete system, configured
Unlimited usage · no per-token fees

Five systems.
Or build your own.

Private AI, sized to your work.

Selected systemNook
R 28 000once-off · hardware includedDownload quote

Hardware and setup included. Delivery extra. Final quote on enquiry.

Enquire about this system

Nook

R 28 000once-off · hardware included
  • Quick drafts and questions
  • Compact desktop
  • For one person
Download quote
What’s inside

Complete computer

Computer
GMKtec NUCBOX M3, Intel Core i5-12450H and Intel UHD graphics
Memory
32 GB DDR4
Storage
1 TB NVMe
Chassis & power
Included mini-PC enclosure, cooling, power adapter and SA mains cord
System
Ubuntu Server LTS, wired LAN and local access controls

Everyday

R 33 000once-off · hardware included
  • Writing and summaries
  • Quiet Mac mini
  • More capable everyday help
Download quote
What’s inside

Complete Mac mini

Computer
Apple Mac mini M6, 12-core CPU and 12-core GPU
Memory
16 GB unified memory
Storage
256 GB SSD
Chassis & power
Apple enclosure, motherboard, cooling, power supply and cord included
System
macOS, a compatible local serving runtime and private LAN access

Desk

R 59 000once-off · hardware included
  • Code and documents
  • Dedicated NVIDIA graphics
  • Fast daily workflows
Download quote
What’s inside

Complete computer

CPU & board
AMD Ryzen 5 5600, Gigabyte B550M K motherboard
Graphics
GeForce RTX 5060 Ti, 16 GB VRAM
Memory
32 GB Corsair Vengeance LPX DDR4 (1 × 32 GB)
Storage
1 TB Crucial E100 NVMe SSD
Power & case
Gigabyte P650G 650 W Gold PSU, Antec VX310 ARGB tower and ID-Cooling SE-214-XT V2 Plus cooler
System
Ubuntu Server LTS, local API and access controls

Think

R 160 000once-off · setup included
  • Larger-model reasoning
  • 128 GB memory · 4 TB storage
  • Lower cost, slower pace
Download quote
What’s inside

Complete AI appliance

Computer
GMKtec EVO-X3, AMD Ryzen AI Max+ 395 with 16-core CPU and Radeon 8060S graphics
Memory
128 GB LPDDR5x-8000 unified memory
Storage
4 TB NVMe SSD
Chassis & power
Factory motherboard, metal enclosure, cooling, power adapter and SA mains cord
System
Linux, a compatible ROCm inference runtime, private LAN endpoint and access controls

Reasoner

R 222 000once-off · hardware included
  • Stronger model reference
  • Faster estimated responses
  • Unlimited local usage
Download quote
What’s inside

Complete AI appliance

Computer
NVIDIA DGX Spark, GB10 Grace Blackwell with 20-core Arm CPU
Memory
128 GB LPDDR5x unified memory
Storage
4 TB NVMe SSD
Chassis & power
Factory enclosure, cooling, 240 W adapter and mains cable
Network
10 GbE, private LAN endpoint and access controls
System
DGX OS, ds4-dfm-rs and private access controls
Model
DeepSeek V4.1 Flash 1017 · calibrated iVQ quant
Performance
30–40 tokens/sec estimated, one active request. Experimental quant; local reasoning quality is unmeasured.

Intelligence power.

Relative intelligence · top tier = 100

Our systemsReference models
  1. LFM2 1.2BReference · est.
    2* ref.
  2. NookR 28 000
    4* ref.
  3. Claude 3 HaikuReference · est.
    13* ref.
  4. EverydayR 33 000
    24 ref.
  5. Claude Opus 4.5Non-reasoning · est.
    53* ref.
  6. DeskR 59 000
    76 ref.
  7. GPT-5.5 HighReference
    82 ref.
  8. ThinkR 160 000
    93 ref.
  9. ReasonerR 222 000
    100 ref.
  10. EnterpriseBuilt around your workload
    Custom scope
Desk: speed and value. Think: lower-cost reasoning. Reasoner: higher model reference. Parent references only; local quality and Reasoner–Think parity are unverified. * Estimated reference. Relative scores do not measure percentage gains in answer quality.
Sources & estimates

Reasoner reference checked 22 October 2026: 45 versus Think’s 42. Other references retain the existing benchmark snapshot. Reasoner’s 30–40 tokens/sec is a planning estimate for calibrated iVQ. A related 1017 VQ quant on one Spark reported 31.5 for prose and 39.4 across mixed tasks, with thinking off. iVQ speed and reasoning quality are unmeasured. One request runs at a time; others queue. Confirm model acceptance and speed for your workload.

Estimated speed.

Desk for speed. Think for value. Reasoner for a higher model reference.

  1. Nook
    10–25 tokens/sec
  2. Everyday
    15–35 tokens/sec
  3. Desk
    40–55 tokens/sec
  4. Think
    12–16 tokens/sec
  5. Reasoner
    30–40 tokens/sec
  6. EnterpriseBuilt around your workload
    Set your target
Estimated tokens/sec for one active stream after warm-up at modest context. Reasoner’s range is inferred from a related quant with thinking off; reasoning speed may differ. Exact builds are unmeasured. Think and Reasoner use experimental runtimes.

Selling price.

Once-off. Hardware, setup and handover included.

  1. Nook
    R 28 000
  2. Everyday
    R 33 000
  3. Desk
    R 59 000
  4. Think
    R 160 000
  5. Reasoner
    R 222 000
  6. EnterpriseBuilt around your workload
    Custom quote
Think offers lower-cost reasoning. Reasoner adds a higher model reference, faster estimated responses and a 4 TB AI appliance. Delivery extra; final scope confirmed on enquiry.

More than
a machine.

Hardware, setup and handover included.

Private by default

Prompts and documents stay inside your network by default.

Ready for your apps

Connect your tools. Unlimited local usage, with no per-token fees.

Tuned to your work

We size and configure the system around your documents, workflow and team.

Hardware, setup and handover.

One package, configured for your team, with access controls and a clear handover.

Keep the work
close.

Tell us your workload, team size and response target.

Enquire on WhatsApp