AI that runs on your hardware,
where your data never leaves.

Open models like Llama, Mistral and Qwen run inside your own network, your own cloud account, or a machine with no internet connection at all. No prompt, no document and no customer record is sent to anyone else. We size it, install it, connect it and keep it running.

This is for you if

Where this is used

Six reasons companies keep the model in house

The engineering is the same in each case. What differs is the reason it has to be private, and that reason decides how strictly it is set up.

Rules

Data that cannot leave the country

Banking, government, health and defence work in the Gulf and in Europe often carries a hard requirement that data stays inside a named border. A model in your own regional cloud satisfies it. Typical driver: data residency clauses, national regulation.

Rules

Client contracts that forbid third parties

Law firms, accountants and engineering consultancies often cannot send a client's material to any outside processor, whatever the terms say. Private removes the question. Typical driver: confidentiality clauses, professional duty.

Cost

High volume where per-question pricing hurts

Once you are processing hundreds of thousands of items a month, a fixed machine cost can beat a per-request bill. We work out where your break-even actually sits before recommending anything. Typical driver: document processing at scale.

Practical

Sites with no reliable internet

Factories, ships, mines, remote plants and secure rooms. The model sits on a machine on site and keeps working when the connection does not. Typical driver: offline or air-gapped operation.

In your product

Software you install at a customer's site

If you sell on-premise software, your AI features have to run there too. We package the model so it ships and updates with your product. Typical driver: on-premise product releases.

Control

Behaviour that must not change underneath you

A hosted model is updated by its provider, and answers can shift without warning. A model you hold changes when you decide to change it, which matters when something has been validated. Typical driver: validated or audited processes.

Services

Everything we do to get it running and keep it there

Installing a model is a morning's work. Making it fast enough, reliable enough and secure enough to put in front of staff or customers is the rest of it.

01

Work out whether you need this at all

Sometimes a hosted model in your own region, under the right agreement, meets the requirement at a fraction of the effort. We check that first and tell you when it is true, even though it is the smaller job for us.

02

Choose the model, by testing rather than by reputation

We run the same set of your real tasks through several open models and compare accuracy, speed and cost. Rankings published online are measured on general knowledge tests, which have almost nothing to do with reading your invoices.

03

Size the hardware honestly

How much graphics memory the model needs, how many people can ask at once, and what happens at your busiest hour. Undersized hardware is the most common reason a private deployment gets abandoned in month two.

04

Make it smaller where that is free

A compressed model can run on much cheaper hardware. Some compression costs you almost no accuracy and some costs a lot. We measure which on your own tasks rather than accepting the general claim.

05

Install and serve it properly

A serving layer that handles many people at once, queues fairly, restarts cleanly, and does not fall over when a request is unusually long. This is ordinary infrastructure work, and skipping it is what makes private deployments feel slow.

06

Put one door in front of every model

Your applications talk to one internal endpoint, not to a specific model. That is what makes swapping a model later a configuration change rather than a rewrite, and it is where logging, limits and access control live.

07

Connect it to your identity and permissions

Single sign-on, groups and roles, so who may use which model is decided in the system you already manage people in. Every request is attributed to a person.

08

Secure it against the attacks specific to AI

Instructions hidden inside documents, attempts to make the model reveal its configuration, and requests designed to run up cost. These are not ordinary web risks and ordinary web testing does not find them.

09

Give your auditors what they will ask for

Where the model runs, what is logged and for how long, who has access, and evidence that no data leaves. Prepared as a document you can hand over rather than assembled in a panic during the audit.

10

Keep it current

Better open models arrive every few months. We retest against your own task set before changing anything, so you upgrade on evidence and can roll back if the new one is worse on the cases you care about.

Tell us your volume and your restriction, and we will tell you what hardware it needs.

The trade

What you gain and what it costs you

Private is not automatically better. It is better for some requirements and worse for others, and you should see both columns before deciding.

What changesModel you run yourselfModel as a hosted service
Where your data goes Nowhere. It stays inside your network To the provider, under their agreement
What you pay for Hardware and running it, whether busy or idle Each request. Nothing when nobody is asking
Raw capability Behind the best hosted models on the hardest reasoning The strongest available, updated constantly
Who controls changes You. It behaves the same until you upgrade it The provider. Behaviour can shift without notice
Effort to set up Weeks, and someone has to own the machine An afternoon
Works with no internet Yes, including fully air-gapped No

Our honest position: if you have no rule forcing your hand and your volume is modest, use a hosted model in your own region. Private earns its cost when a rule, a contract or your volume makes it the cheaper answer.

Our stack

The tools, standards and methods we use

Listed so your infrastructure and security teams can see exactly what would be installed, and confirm that none of it is exotic.

The models themselvesOpen weight models you are licensed to run and keep. We check the licence covers commercial use before recommending one
LlamaMistralQwenGemmaDeepSeekPhi
What serves the modelHandles many people at once and keeps the graphics card busy, which is what makes it feel fast rather than sluggish
vLLMText Generation InferenceOllamallama.cppTensorRT-LLM
Making it smallerCompression lets a large model run on cheaper hardware. We measure the accuracy cost on your own tasks first
GPTQAWQGGUF4-bit and 8-bitSpeculative decoding
The hardwareSized from your busiest hour, not your average. We tell you the cheapest card that will actually cope
NVIDIA H100A100L40SRTX 6000 AdaApple silicon for small models
Where it runsYour own servers, your own cloud account, or a sovereign region. Air-gapped where the rules require it
KubernetesDockerAzure and AWS private regionsBare metalAir-gapped installs
The single door in front of itOne internal endpoint your applications call, so the model underneath can change without touching them
OpenAI-compatible APILiteLLMModel gatewayRate limitsPer-team quotas
SecurityIncludes the risks specific to AI systems, which ordinary application testing does not look for
OWASP Top 10 for LLMsPrompt injection testingSingle sign-onNetwork isolationSecrets management
Watching itSpeed, queue length, card usage and cost per thousand requests, so you can see what you are getting for the hardware
PrometheusGrafanaTime to first wordWords per secondGPU utilisation

Example

When private turned out to be the wrong answer

Example

A professional services firm came to us certain they needed everything on their own hardware. Their client agreements forbade sending client material to third parties, and that reading was correct.

When we measured the actual work, it came to a few thousand documents a month. The hardware to serve that reliably, with a spare machine so a failure did not stop the firm, cost more per year than the hosted equivalent by a wide margin, and needed somebody to look after it.

What actually satisfied their obligation was a hosted model in their own region under an agreement with no training on their data and no retention, plus written confirmation of where it ran. Their compliance officer accepted that in a week. We built the retrieval and the application, and none of it needed a graphics card.

Two years later, if their document volume grows tenfold, private becomes the cheaper option and the single endpoint we put in front of the model means switching is a configuration change. That is why it is there.

An example, with invented figures, to show how the decision is actually made. Your volumes and obligations will point somewhere else.

FAQ

Questions we get asked before starting

Are open models good enough?

For reading documents, pulling out fields, classifying, summarising and answering from your own material, yes, comfortably. For the hardest open-ended reasoning the best hosted models are still ahead. Most business work is in the first group, and we test on your tasks rather than argue about it.

What does the hardware cost?

It depends entirely on model size and how many people ask at once, and we will not quote a figure before measuring your load. What we can do quickly is tell you which class of card you are looking at, which is usually the answer you actually need.

Can it run with no internet at all?

Yes. Fully air-gapped installations are a normal request, particularly in defence and critical infrastructure. Updates are then a deliberate, scheduled act rather than something that happens quietly.

Who looks after it once it is running?

Your team, ours, or both. If you would rather not own it, this is covered by our managed operations work, which includes watching it, patching it and testing model upgrades before they go anywhere near your users.

Can we start hosted and move later?

Yes, and it is often the right order. Because everything talks to one internal endpoint rather than to a specific provider, moving later is a configuration change instead of a rebuild. We set it up that way from day one whichever you choose.

Does this satisfy the EU AI Act?

Running privately helps with the data protection side but it is not the whole answer. The Act asks about risk classification, human oversight and record keeping, which are about how you use the system rather than where it sits. We cover both when the work needs it.

Next step

Tell us the rule you have to satisfy.

The regulation, the contract clause, or the volume that made you look at this. We will tell you on the call whether private is genuinely required, and if it is, roughly what it takes to run.

Ahmedabad, India. We work with teams in the US, UK, Europe, Singapore and the Gulf, and we are used to the time difference.