Cape CodesEnquire

Your AI.
On your ground.

Run useful AI on your own network, with the computer, model and software set up for your team.

Explore the systems
On premiseLocal inference / 01Private by design
One private boundary.Models, prompts and responses stay on your network by default.
10+tokens / sec
target
01 / Hardware02 / Model03 / Serving04 / Validation
Private on your network
Complete system, configured
10+ tokens/sec target

Five systems.
Sized to the work.

Choose a workflow. We’ll size the model and memory, then verify speed on your finished system.

Selected systemNook
R 20 000once-off · configured system

Includes the system, setup and model configuration. Final parts are confirmed against stock and workload.

Enquire about this system
01 / 05One person

Nook

Compact, quiet AI for focused work.

R 20 000Once-off · complete system

16 GB memory · 512 GB SSD · small Q4 models

See complete hardware & model fit

Complete computer

Computer
Intel Core i5-12450H mini PC with Intel UHD graphics
Memory
16 GB DDR4-3200
Storage
512 GB M.2 SSD
Chassis & power
Integrated enclosure, cooling and power adapter
System
Ubuntu Server LTS, wired LAN and local access controls

Model shortlist

Qwen3 0.6B Q4_K_M and 1.7B Q4_K_M for light, focused tasks.

Ollamallama.cppGGUF

Validated for one active request on a small Q4 model.

02 / 05Small team

Desk

A first GPU for shared office tasks.

R 40 000Once-off · complete system

RTX 3060 · 12 GB VRAM · Qwen3 8B Q4

See complete hardware & model fit

Complete computer

CPU & board
AMD Ryzen 5 5500 with boxed cooler, A520 micro-ATX motherboard
Graphics
Tested GeForce RTX 3060, 12 GB VRAM
Memory
32 GB DDR4-3200
Storage
1 TB NVMe SSD
Power & case
650 W 80+ Gold PSU, mesh-front micro-ATX case with two included fans
System
Ubuntu Server LTS, local API and access controls

Model shortlist

Qwen3 8B Q4_K_M; 14B Q4_K_M at a shorter context; Phi-4-mini Q5_K_M.

Ollamallama.cppGGUF

Best suited to one 8B model and a small local team.

03 / 05Model workstation

Studio

More memory for capable local models.

R 70 000Once-off · complete system

RTX 5060 Ti · 16 GB VRAM · Qwen3 14B Q4

See complete hardware & model fit

Complete computer

CPU & board
AMD Ryzen 5 5600, B550 micro-ATX motherboard, boxed cooler
Graphics
GeForce RTX 5060 Ti, 16 GB VRAM
Memory
32 GB DDR4-3200
Storage
1 TB NVMe SSD
Power & case
650 W 80+ Gold PSU, mesh-front micro-ATX case with two fans
System
Ubuntu Server LTS, model storage, local API and access controls

Model shortlist

Qwen3 14B Q4_K_M; 8B Q5_K_M; Phi-4-mini Q5_K_M.

Ollamallama.cppGGUFvLLM*

We tune context to fit the model in GPU memory.

04 / 05Growing team

Workgroup

Two systems for parallel team workflows.

R 150 000Once-off · two configured systems

Two RTX 5060 Ti systems · 16 GB VRAM each · shared API

See complete hardware & model fit

Two complete computers

Each CPU & board
AMD Ryzen 5 5600, B550 micro-ATX motherboard, boxed cooler
Each graphics card
GeForce RTX 5060 Ti, 16 GB VRAM
Each memory & storage
32 GB DDR4-3200 and 1 TB NVMe SSD
Each power & case
650 W 80+ Gold PSU, mesh-front micro-ATX case with two fans
Shared setup
Private model API, simple health dashboard and access controls

Model shortlist

Run separate copies of Qwen3 14B Q4_K_M, or give each node a separate workflow.

Ollamallama.cppPrivate API

Each node handles separate work in parallel.

Memory sets
the model size.

Model fit depends on memory, quantisation and context. We test each choice on the finished system.

12 GB · value GPU

RTX 3060

A lower-cost CUDA option for smaller local models.

Qwen3 8B Q414B with shorter context
16 GB · current midrange

RTX 5060 Ti or RX 9060 XT

16 GB of VRAM for larger models at practical power use.

14B Q4Ollama / llama.cppCheck exact backend
24 GB · specialist GPU

RTX 3090

A tested used card opens a 24 GB path to stronger models.

Qwen3.8 27B Q4Mistral Small 24B Q4Quoted to stock
128 GB · unified memory

DGX Spark

128 GB unified memory for larger models. Fit and speed depend on context.

Ollamallama.cppSGLang

Benchmarks are a starting point.

We test candidate models on your prompts before recommending one.

View a model comparison
MoE models still need room for all weights.

Active parameters describe compute per token, not total memory. Plan for full weights, context cache and runtime.

Read the model card

Choose a model
and quant.

Quantisation trades memory use against quality. We test the best fit for your prompts.

Serving software

Ollama for simple local APIs, llama.cpp for lean GGUF inference, and vLLM for supported GPUs and concurrent requests.

Ollama llama.cpp vLLM

DGX Spark supports models up to 200B parameters in some workflows; fit and speed vary. We confirm model, runtime and licence before delivery. DGX Spark specifications

Built, tuned
and handed over.

We set up the computer, model and private endpoint as one system.

01

Choose a workflow

Agree on prompts, documents and users.

02

Build and configure

Set up the OS, model, quant and private API.

03

Verify performance

Measure speed on the finished system.

10+ tokens per second

Target 10+ generated tokens/sec on one active stream after warm-up, using the agreed model and prompts. Longer context and concurrent use can change speed.

Keep the work
close.

Tell us your use case and we’ll size a system around it.

Enquire on WhatsApp