Our current turnaround time is 7-10 working days
What are you building for?
Tailored recommendationsPC builds across the site will be filtered to match your chosen use-case.

LLM Workstations

Run Llama 3, Mistral and 70B-class LLMs locally.

Hand-built inference and fine-tuning workstations with the VRAM, system memory and storage throughput to actually keep up with modern LLM workflows. From single-RTX-5090 dev rigs to dual RTX 6000 Pro Blackwell production stations.

  • Up to 192GB VRAM (dual RTX 6000 Pro)
  • Pre-tested with Ollama, vLLM, llama.cpp
  • Optional onboarding session
  • 5-year warranty
Configure NowTalk to a Specialist

Recommended configurations

Three calibrated starting points. Every component is configurable in our builder.

Most popular

LLM Developer

Local inference up to 70B (Q4)

£3,499

Starting price

CPU
Intel Core Ultra 9 285K
GPU
NVIDIA RTX 5090 (32GB)
RAM
64GB DDR5-6400
Storage
2TB Gen5 NVMe
Best for
Solo dev, RAG, AnythingLLM, Ollama
Configure LLM Developer

LLM Studio

Fine-tuning + inference, single Pro GPU

£8,499

Starting price

CPU
AMD Threadripper PRO 7965WX
GPU
NVIDIA RTX 6000 Pro Blackwell (96GB)
RAM
128GB DDR5 ECC
Storage
4TB Gen5 NVMe + 8TB NVMe
Best for
Fine-tune up to 13B · serve 70B at scale
Configure LLM Studio

LLM Production

Dual Pro GPUs, 192GB VRAM pool

£16,999

Starting price

CPU
AMD Threadripper PRO 7995WX
GPU
2× NVIDIA RTX 6000 Pro Blackwell
RAM
256GB DDR5 ECC
Storage
4TB Gen5 NVMe + 16TB NVMe
Best for
180B+ inference, 70B fine-tune
Configure LLM Production

Which model fits which workstation?

Sizing local LLMs is mostly a VRAM problem. The table below assumes 4-bit quantisation with a 4K–8K context budget. Drop to 8-bit for higher quality and double the VRAM requirement; double again for FP16 / training.

Which model fits which workstation?
 Parameter countQ4 VRAM (approx.)Recommended GPUCREATE tier
Llama 3.1 8B
~6GBRTX 5060 Ti 16GB+Developer (entry)
Mistral / Llama 13B
~10GBRTX 5070 Ti / 5080Developer
Mixtral 8×7B
~28GBRTX 5090 (32GB)Developer
Llama 3.1 70B
~40GBRTX 5090 w/ offload or RTX 6000 ProDeveloper / Studio
DeepSeek 67B
~40GBRTX 6000 Pro Blackwell (96GB)Studio
Llama 3.1 405B
~220GB2× RTX 6000 Pro + offloadProduction

Throughput depends on framework (vLLM / llama.cpp / Ollama), quantisation (GGUF / GPTQ / AWQ) and context length. Figures are practical guidance — we benchmark each build with your chosen model before dispatch.

Why CREATE PCs

Tuned for sustained GPU load

PSU, AIO and case airflow specified to hold full GPU power without thermal throttling on long generations.

PCIe-aware multi-GPU builds

When you need two cards, we use boards with proper x8/x8 lane splitting and chassis that fit dual triple-slot GPUs with airflow.

Onboarding included on Studio+

A real conversation about your stack — Ollama, vLLM, LM Studio, RAG — and we set up CUDA, cuDNN and drivers before delivery.

LLM Workstations FAQs

Can I really run Llama 3.1 70B on a single RTX 5090?

Yes, with 4-bit quantisation (GGUF Q4_K_M) using llama.cpp or Ollama, with about 8K context. Speeds are typically 12–22 tok/s. For long context (32K+) or higher quality (Q5/Q8) you want an RTX 6000 Pro Blackwell with 96GB VRAM.

Should I buy two RTX 5090s instead of one RTX 6000 Pro?

For pure inference of models that fit in 32GB, two RTX 5090s can be cheaper than one RTX 6000 Pro and offer parallel inference. But Blackwell consumer cards do not support NVLink, so you cannot pool VRAM — 2×32GB is not the same as one 64GB card for big models. For models that need 40–96GB VRAM in one process, choose the RTX 6000 Pro. We can build both — speak to us.

Do you preinstall Ollama / vLLM / LM Studio?

On request, yes. We install Windows 11 Pro by default with NVIDIA Studio drivers and CUDA. We can preinstall Ollama, LM Studio, vLLM under WSL2, AnythingLLM and your chosen model weights before delivery — just specify at order.

What about ECC RAM for training?

We strongly recommend ECC RAM for fine-tuning runs longer than ~12 hours. The LLM Studio and LLM Production tiers ship with ECC by default; the Developer tier can be upgraded.

Linux or Windows?

Either. Most clients want Windows 11 Pro with WSL2 Ubuntu for the AI stack — that is what we ship by default. Pure Ubuntu LTS is available on request with the same warranty.