Hardware, built & shipped
From a silent Mac mini to multi‑GPU RTX 5090 workstations and RTX PRO 6000 Blackwell nodes — specced, assembled and burned‑in for inference.
On‑premise AI · your data stays in your building
Local AI hardware and software, built to run open LLMs on‑prem.
Pico Software designs, builds and configures on‑premise hardware — and installs the full local‑LLM software stack — so companies and individuals can run today's best open‑weight models in‑house. No cloud. No prompts or documents leaving your network.
Apache 2.0 / MIT open models
Data never leaves your network
UK build, install & support
On‑prem inference node
In your building · air‑gapped capable
Memory in use
Llama 3.3 70B running
Qwen3‑235B‑A22B fits @4‑bit
GLM‑4.5 · 355B fits @4‑bit
Running locally · 0 bytes sent to cloud
Hardware + software, end to end
Everything you need to run AI in‑house
We supply the box and the brains: the right on‑prem hardware, the local‑LLM software stack, and the help to get it running for your team.
Local‑LLM software stack
We install and tune the inference stack — open‑weight models, runtime, OpenAI‑compatible API and a chat UI — so your team can chat, code and build on day one.
Setup, tuning & support
On‑site or remote install, model selection, quantisation and ongoing support — sized to your workloads, context lengths and budget.
Total data sovereignty
Everything runs inside your building. Prompts, documents and model weights stay on your network — ideal for regulated and privacy‑first teams.
Hardware tiers & pricing
Pick the box that fits your models
Memory capacity decides which models fit; bandwidth decides how fast they generate. Each tier lists the largest open‑weight models that fit in memory at 4‑bit.
Mac Mini
Compact & silent
Apple M4 / M4 Pro
Up to 48GB unified memory
Approx. cost
£800–£1,400
single box · entry on‑prem
Get a quote
Largest models it runs (4‑bit)
- Llama 3.3 70B 70B
- Mixtral 8×7B 46.7B
- Qwen3‑32B 32B
Most popular
Single‑or‑Multi GPU
Scalable NVIDIA workstation
1–4× RTX 5090
32–128GB VRAM
Approx. cost
£2,800–£10,600
1–4 GPUs · best £/perf
Get a quote
Largest models it runs (4‑bit)
- Qwen3‑235B‑A22B 235B
- Command A+ 218B
- Mixtral 8×22B 141B
RTX PRO 6000 Blackwell
Prosumer flagship
1–2× RTX PRO 6000 Blackwell
96–192GB ECC GDDR7
Approx. cost
£8,000–£18,000
1–2 cards · dept. server
Get a quote
Largest models it runs (4‑bit)
- GLM‑4.5 355B
- Qwen3‑235B‑A22B 235B
- Command A+ 218B
Indicative UK pricing (mid‑2026) and example open‑weight models — the largest that fit in memory at 4‑bit quantisation, where all parameters must fit in RAM/VRAM (Mixture‑of‑Experts models keep every parameter resident even though only a subset is active per token). Prices were moving month‑to‑month under a DRAM shortage; figures are from our June 2026 on‑prem LLM hardware report — please re‑check before ordering.
Why on‑premise
Keep the model — and the data — in the building
-
1
Your data never leaves your network
Run inference entirely on‑prem — no prompts, files or embeddings sent to third‑party APIs.
-
2
Right‑sized to the models you need
Memory capacity decides which models fit; we match the hardware to your target models and context lengths.
-
3
Open, commercially‑usable weights
We focus on Apache 2.0 / MIT models you can deploy and build on commercially — and we flag the licensing traps.
-
4
Built, installed & supported in the UK
Assembly, burn‑in, on‑site install and ongoing support — whether it's one desktop or a department server.
On‑prem, by the numbers
0bytes of your data sent to the cloud during inference.
70–355Bparameter open models running on a single workstation.
4‑bitquantisation fits big models into a fraction of the memory.
UKdesign, build, install and support — start to finish.
Bring AI in‑house.
Tell us your use case and target models — we'll spec the right on‑prem build and quote it.
sales@pico-software.com