# Private LLM Hosting UK

Source: https://dijihost.co.uk/private-llm-hosting/
Updated: 2026-11-12

> dijitul hosts private large language models for UK businesses, quoted to the model and your usage. We run open-weight models with Ollama on a GPU server, or on CPU for small models, lock the API behind authentication and HTTPS, and manage the server. If an API would serve you better, we will say so.

## Key facts

- Quoted per server, sized to the model you want to run
- Ollama (MIT licence) serving open-weight models through an HTTP API
- GPU servers for real-time use; CPU only for small models or batch jobs
- Ollama binds to 127.0.0.1:11434 by default; we keep it private behind a proxy
- Authentication and HTTPS added in front, because the docs describe no built-in auth
- Each model has its own licence, which you need to check
- Prompts and outputs stay on the server you control

## What "private LLM" means

A private LLM is an open-weight language model that runs on a server you control, instead of sending prompts to a model provider. People choose it for three reasons: data must not leave their own infrastructure, they have steady high volumes where a fixed server cost works out cheaper, or they need to run offline or with a fine-tuned model.

We usually use **Ollama**, which is MIT licensed and serves open-weight models such as Qwen, Gemma, DeepSeek and gpt-oss through a simple HTTP API. Your app, or tools like Open WebUI, talk to that API. For larger multi-user setups we can also run other inference servers.

## GPU or CPU?

| | GPU server | CPU-only server |
| Speed | Fast enough for chat and live apps | Much slower per answer |
| Model size | Limited by GPU memory (VRAM) | Limited by system RAM |
| Best for | Chatbots, assistants, many users | Small models, embeddings, overnight batch jobs |
| Cost | Higher fixed monthly cost | Lower |

Ollama's GPU docs list NVIDIA cards with compute capability 5.0 or higher on driver 550 or newer, AMD cards through ROCm v7, Apple Metal, and Vulkan on Windows and Linux. Memory use grows with model size, quantisation, context length and parallel requests: Ollama's FAQ notes that memory scales with OLLAMA_NUM_PARALLEL multiplied by the context length. We size the server from the exact model and settings you plan to use. See [GPU servers](https://dijihost.co.uk/gpu-servers/) for the hardware side.

## Keeping it private and secure

- Ollama listens on 127.0.0.1 port 11434 by default. We leave it there and publish it only through an Nginx reverse proxy with HTTPS.
- Ollama's documentation does not describe built-in authentication, so we add it at the proxy: API keys, single sign-on or an IP allow-list.
- The firewall blocks everything else, and the server is patched and monitored.
- Logs record usage, not prompt content, unless you ask for prompts to be logged.
- Model files are stored on fast disk and included in the backup plan, or re-downloaded if that is quicker.

## Settings that matter in practice

- **Context length**: Ollama's default is 4,096 tokens. Long documents need more, which costs memory.
- **Parallel requests**: the default is one request per model at a time. More users need a higher setting and more memory.
- **Keep-alive**: models stay loaded for 5 minutes after use by default, so the first request after a quiet spell is slower. We can keep a model loaded all the time.
- **Model licences**: each open-weight model has its own licence and acceptable use terms. Check them before you use a model commercially.

*Sources, checked 11 November 2026: [Ollama FAQ](https://docs.ollama.com/faq); [Ollama GPU support](https://docs.ollama.com/gpu); [Ollama on GitHub (MIT licence)](https://github.com/ollama/ollama).*

## When an API is the better choice

Self-hosting is not always right. If you want the strongest models, low volumes, or no hardware to look after, a hosted API is usually cheaper and better. Frontier models are not available as open weights, and a GPU server costs the same whether it is busy or idle. If UK data residency is the concern, our sister service [dijitul.ai](https://dijitul.ai) is a UK-hosted API gateway for Claude with inference in London. Many projects mix the two: an API for answers and a small self-hosted model for embeddings or sensitive steps. For apps built on top of a model, see [LLM app hosting](https://dijihost.co.uk/llm-app-hosting/).

## FAQs

### Can I run my own LLM on a UK server?

Yes. dijitul runs open-weight models with Ollama on managed UK GPU or CPU servers, secured behind HTTPS and authentication. It is quoted to the model size and how many people will use it.

### Do I need a GPU to run an LLM?

Not always. Small models can run on CPU, but answers are much slower, which suits batch jobs more than live chat. For chatbots or many users, a GPU server is the practical choice. The model size and context length decide how much GPU memory you need.

### Is Ollama secure to expose to the internet?

Not on its own. Ollama binds to 127.0.0.1:11434 by default and its documentation does not describe built-in authentication. Put it behind a reverse proxy with HTTPS, authentication and a firewall. dijitul sets this up on servers it manages.

### Is a private LLM cheaper than an API?

Only at steady, high volumes. A GPU server is a fixed monthly cost whether it is busy or idle, while an API charges per token. For low or unpredictable use, or when you need the best model quality, an API is usually cheaper.

### Can I use open-weight models commercially?

Often, but it depends on the model. Each open-weight model has its own licence and acceptable use policy, and some have restrictions. Ollama itself is MIT licensed, but that does not change the licence of the models you download through it.

### How much memory does an LLM need?

It depends on the parameter count, the quantisation, the context length and how many requests run in parallel. Ollama notes memory scales with parallel requests multiplied by context length. dijitul sizes servers from the exact model and settings you plan to use.

## Pricing and ordering

This is quoted to fit. Request a quote at https://dijihost.co.uk/quote/ or call 01623 650333.
