Private AI
The model runs on your hardware. Nothing leaves the building.
Most AI tools send your records, contracts and case files to someone else's servers and bill you per word. We deploy open-weight models inside your network, fine-tune them on your documents, and install the whole thing on site. Twelve years of shipping production systems, not a reseller with a markup.
/// the blocker
You are not slow to adopt AI. You are not allowed to leak data.
Hospitals, law firms, manufacturers under NDA, financial services, HR departments. The appetite for AI is already there. What kills the project is a data processing agreement nobody will sign, or a client contract saying your files stay on your systems.
A private deployment removes that conversation entirely. The model runs where the data already lives, so there is no third party to vet, no retention policy to negotiate, and no vendor who can change the terms underneath you six months in.
/// what we deploy
Five pieces. One private stack.
Most organisations need three of these, not all five. We scope it to the workflow that is actually costing you hours.
Model Serving
An open-weight model served inside your network behind an OpenAI-compatible endpoint, so the applications you already have can talk to it without a rewrite. vLLM or SGLang, sized against your real concurrency rather than a guess.
Fine-Tuning On Your Corpus
LoRA or full-parameter tuning on your own documents, so the model knows your formats, terminology and edge cases. Shorter prompts, output your systems can parse, and a training set that stays on your storage.
Retrieval Over Your Documents
A vector index that lives next to the model, respects your existing permission model, and cites the source document in every answer. Nothing is uploaded to an outside index.
Agents Wired Into Your Systems
The part that actually pays. A model on its own answers questions. Agents connected to your HIS, case management, ERP or ticketing do the work. Tool use, browser control and computer use all run locally.
Evaluation And Guardrails
A private model with no evaluation harness is a liability. We build the reference set with your experts and score every model change against it, so you find out about quality drift before your users do.
/// hardware, honestly
What the current open models actually need.
Vendor minimums assume aggressive offloading and describe the smallest model in the family. Where secondary sources disagreed by a factor of three, we went to the official repository. These are the video generation models, which are the most demanding class anyone asks us about. Text models are considerably cheaper to run.
The real constraint
VRAM decides which models you can load. Card count decides how much work you get through. These models parallelise across GPUs without NVLink, so once you clear the memory bar, buying more cards buys throughput almost linearly.
| Model | Params | Vendor floor | Production reality | Licence |
|---|---|---|---|---|
| LTX-2.5Lightricks | 22B | 16 GB | 48 to 96 GB for 4K with headroom | LTX-2.x Community, free under $10M ARR |
| Wan 2.2 T2V-A14BAlibaba | 27B MoE, 14B active | 80 GB | 80 GB, H100 or a 96 GB workstation card | Apache 2.0 |
| Wan 2.2 TI2V-5BAlibaba | 5B dense | 24 GB | 24 to 32 GB | Apache 2.0 |
| HunyuanVideo 1.5Tencent | 8.3B | 14 GB | 24 to 48 GB | Tencent community terms |
| HunyuanVideo (original)Tencent | 13B | 60 GB | 100 to 120 GB at 1080p, so an H200 | Tencent Community Licence |
The 16 GB floor is marketing
Wan 2.2's own repository says 80 GB minimum for the 14B model, with offloading already switched on. Sites quoting 24 GB are describing the 5B model.
Headline speeds hide the hardware
LTX-2.5's 6.8 seconds for a 10-second clip was measured on two GB200 superchips. Through the vendor's own API the same job takes 23.7 seconds.
Licences are not all Apache
Only the Wan family is plainly permissive. LTX-2.5 is free below $10M revenue and negotiated above it. Tencent uses its own terms. This decides what you can build on.
Generating video rather than text? The economics are completely different, and so is the buying decision. We built a separate page with the hardware tiers, the throughput maths and a cost-per-delivered-clip calculator. Private AI for video production →
/// how the install works
Our engineers come to you. Nothing dials home.
There is an easy version of this where we install remote access software and work from another country. We do not do that, because it hands you back the risk you are paying to remove.
Built on site
Our engineers work in your building, on your network, on the server you own. Model, fine-tuning and integrations all happen behind your firewall while your team watches it being done.
No access left behind
When we leave there is no backdoor, no VPN tunnel to us and no agent phoning out. The server can sit fully air-gapped. Your IT team holds every credential.
Kept current on your terms
When a better open model lands we test it against your reference set first, take a full backup, then an engineer applies it on site. The model never holds delete permissions.
/// where it runs
Three ways to keep it private.
Private does not have to mean a server room. It means your data is processed inside a boundary you control. We will tell you which of these fits your policy and your budget.
On-premise hardware
A GPU server in your own rack. Nothing touches the public internet, which is the only configuration that satisfies a genuinely air-gapped policy.
Best for Strict policy, steady volume
Your private cloud
The model runs inside your own AWS, Azure or GCP account, in your VPC, under the security controls you already operate. No capital spend and no hardware to look after.
Best for Fast start, no hardware
Hybrid routing
Sensitive requests go to the local model, everything harmless keeps using a hosted API. The cheapest way to get most of the compliance benefit without moving your whole stack.
Best for Mixed workloads
/// is this for you?
Worth it for some. Overkill for others.
This fits if
- ✓You handle data under NDA, HIPAA, GDPR or client confidentiality
- ✓Legal or security has already blocked a public AI tool
- ✓You have repetitive document, support or analysis work worth automating
- ✓Your API spend is climbing and shows no sign of stopping
- ✓You want a system that survives a vendor changing its terms
This does not fit if
- ✗You are comfortable sending your data to a hosted API
- ✗Your volume is low enough that per-token pricing is simply cheaper
- ✗You need frontier-level reasoning on open-ended problems today
- ✗You want it live next week with no evaluation step
- ✗Nobody internally will own the system after handover
We benchmark before you spend anything on hardware. If a private model cannot reach the quality bar on your use case, we will tell you plainly and point you at the cheaper answer, even when that answer is not us.
/// questions
Before you book.
How much does a private AI deployment cost?+
Request a scoped assessment covering the workload, data, deployment environment and integrations. Hardware or private-cloud capacity, model evaluation, fine-tuning and installation affect the budget. Separate the engineering work from hardware, electricity, hosting and any ongoing care plan when comparing options.
How long does a private AI deployment take?+
Timing depends on data preparation, model evaluation, infrastructure availability and the systems being connected. Plan the assessment and benchmark before hardware procurement or installation. Network access arrangements, on-site scheduling and acceptance testing also need to be included in the project scope.
Is an open model good enough to replace GPT or Claude?+
For narrow, well-defined work such as extraction, classification, summarising your documents or drafting in your house format, a fine-tuned open model is usually indistinguishable and far cheaper to run. For wide open-ended reasoning and for frontier-level coding, the hosted models are still ahead. We benchmark the shortlist on your actual tasks and show you the scores before you buy hardware.
What hardware will we need?+
For text models, more workloads than people expect fit on a single 48 to 96 GB card. The number depends on model size, context length and how many people hit it at once. We size it during the assessment and give you the figure before you spend anything. For video generation the numbers are much larger, and we publish them on the video page.
Do we still pay per token once it is installed?+
No. The model runs on hardware you own, so no provider meters it. Your running cost is electricity, cooling and an optional care plan. It does not change whether you send ten requests a day or ten thousand.
How do you install it without remote access to our network?+
Our engineers travel to your site and do the work there. We do not install remote access software, because that would reintroduce the exact risk you are paying to remove. When we leave there is no tunnel back to us and your team holds every credential.
Who owns the model afterwards?+
You do. Weights, fine-tuned checkpoints, training data and deployment configuration all sit on your storage. If you stop working with us, nothing switches off and any competent engineer can take it over.
Our staff do not work in English. Does that break it?+
No. These models reason best in English, so that is what runs underneath, and output comes back in your team's language. Where the interface, the documents or the training data need local expertise we bring in a language partner alongside our engineers.
Find out if private AI is worth it for you.
Bring us the workflow and the constraint. We will tell you which model class handles it, roughly what hardware it takes, and whether a private deployment actually beats an API.