Private AI on servers you control, tuned to run fast and cost less
We run open-weight models, the AI models you can download and run yourself, on hardware you control. One gateway gives every user and app its own key and quota. We tune for speed and cost on your tasks, then hand the system to your team.
For whomFor healthcare, finance, legal and other regulated teams whose data must stay inside their own network, and for any team that wants AI costs it can predict.
- Your staff and apps
- One gateway, a key for each user
- Language models
- Image and video models
- Your documents, indexed inside
Every request goes through one gateway on servers you control
Private AI means running the models on servers you control instead of a provider’s cloud. We install open-weight models from the Llama, Qwen, Kimi and DeepSeek families. They run on your premises, in your cloud region, on our private GPUs, or on a network cut off from the internet (air-gapped). We test them on your own tasks, size them for your load, and hand them to your engineers.
We test models on your tasks before you buy hardware
Test first
We run candidate models on a sample of your own tasks. We measure the quality of each answer, the response time and the cost.
Deployment design
We draw the serving software, the hardware size and the network boundary. Your security team reviews the drawing before you buy anything.
Build
The models and their serving engine, search over your documents indexed inside your network, and the gateway with its keys, quotas and audit log.
Handover
Your engineers get the operating guides and an upgrade plan, and they run the system from then on.
Your security team reviews the design, and you approve it, before any hardware is bought.
Example
A legal team searches its contracts, and the text stays on your network
A hypothetical case, to show how the work runs.
- The legal team wants to ask questions of the contract archive on its file server.
- The contracts are indexed on a GPU server inside your network.
- A lawyer asks which contracts renew this quarter. The question reaches the model through the gateway, under that lawyer’s key.
- The answer lists each contract with its source. The gateway logs the request against the legal team’s quota.
- When a lawyer leaves, IT switches off that key in the gateway. The log keeps every request it made.
Everything reaches the models through one gateway
Five layers, from the servers at the base to the tools your people use. The same rules apply to every user and every app.
What your people use
- Chat workspace
- AI agents
- Image and video studio (ComfyUI)
- Workflows
- Your apps, by API
Gateway
- A key per user and app
- Limited to some models if needed, and switched off at once when someone leaves.
- Quotas and rate limits
- One heavy user does not slow everyone else down.
- Usage and cost logs
- Who used which model, when and how much, with the cost for each team.
- Models change behind it
- Apps talk to the gateway, so a better model can replace the old one without changing them.
Serving engine
- Optimised for your hardware
- Quantised models, batching
- Fast answers for many users
Models
- Language models
- Image and video models
- Search over your documents
Servers you control
- Your premises, your cloud region or our private GPUs
- GPU servers sized to your users
- Encrypted, monitored, backed up
How it runs fast
A large open model can answer quickly on a single server
- Sized to your users. We count how many people will ask at once, how long their documents are and how quickly they need an answer. Memory decides which models fit, so the server is chosen from that count.
- Quantisation. The model runs at lower numerical precision, so it fits in less memory and answers faster. We keep a lower precision only where the tests show the quality holds on your tasks.
- A good serving engine. The engine groups requests together and reuses work it has already done, so many users can share one server without long waits.
- Measured before you commit. We measure speed and answer quality on a sample of your own tasks, on the setup we propose. You decide on those results, before any hardware is bought.
Your engineers run the models after handover
How we work together
A fixed piece of work that starts with the tests, so the decision rests on evidence from your own tasks.
- Models running on hardware you control
- The tests that chose them, with response time and cost measured on your work
- Design documents reviewed by your security team
- Safeguards, monitoring and the audit log compliance requires
- A team that can run it without depending on us
Questions buyers ask
What is an on-premises LLM?
A large language model that runs on servers inside your own network or data centre instead of a provider’s cloud. Your questions, documents and answers stay on your own systems.
Are open-weight models good enough?
For many tasks, yes, and for some, not yet. That is why we test on your own material before any hardware is bought. We report the result even when it favours a hosted model from a large provider.
Start by testing models on your own tasks
Tell us what the data is and where it must stay. Our first reply says what testing on your own tasks would involve.
Start a project