Private LLMs · deployed on site


1xNEO trains a frontier-class language model on your own documents, protocols and history, then racks it inside your office. Complete privacy, full company context, and capability that competes with the best public systems.

0%

on your hardware

0

data sent externally

0 wks

audit to handover

no data leaves the building frontier-class performance trained on your archives runs on one office server no per-token billing audit-ready by design yours to keep permanently no data leaves the building frontier-class performance trained on your archives runs on one office server no per-token billing audit-ready by design yours to keep permanently

Why 1xNEO

Three things public AI cannot give your enterprise

Generic assistants know the internet. Yours should know your company, and it should do so without handing your most sensitive data to someone else's servers.

Privacy

Nothing leaves your office

Inference runs on hardware inside your building. No cloud tenancy, no third-party API calls, no document ever crossing your firewall. Compliance teams point at a machine, not a data-processing agreement.

Full context

It knows your company

Fine-tuned on your contracts, tickets, manuals, transcripts and operating protocols. The model carries the institutional knowledge new hires take years to absorb: your people, your processes, your history.

Quality

Frontier-class capability

Built on open-weight foundation models competitive with the best public systems, then specialized for your domain. Your evaluation sets are written by your own experts, not by us, so quality is measured against your standards.

Software layer

The software

A model shaped by your knowledge, integrated into your stack

We fine-tune open-weight foundation models on your corpus, wrap them in a retrieval layer that connects every internal source, and expose them through APIs your existing tools already speak. Then we keep it current.

Custom fine-tuning

Open-weight base models adapted to your corpus, vocabulary and decision style.

Full knowledge integration

Connects document stores, wikis, ticketing and databases into one retrieval layer.

Workflow integration

REST and gRPC endpoints drop into your existing tools and internal apps.

Continuous retraining

Quarterly retraining keeps the model current as your document base grows.

The hardware

Private infrastructure, racked and labelled in your office

Most teams of 20 to 150 run comfortably on a single GPU workstation in a locked closet. Larger deployments get a two-node cluster with failover. We size it after the data audit, never before, so you buy exactly the capacity your workload needs.

Single-node or clustered

Sized from your real query volume and document count.

Air-gap capable

Runs fully offline; updates delivered by encrypted media if you prefer.

Reference specification

Compute

4×–8× enterprise GPU

per inference node

Memory

384–1024 GB

system + VRAM

Storage

NVMe array

redundant, encrypted

Footprint

4U–8U rack

locked closet or cabinet

Network

Air-gap capable

offline updates by media

Availability

99.9% uptime

cluster failover option

Use cases

Built for teams with knowledge worth protecting

Legal & compliance

Reads decades of case files and precedents; drafts stay inside the firm.

Financial services

Answers from internal policy and risk memos without exposing client data.

Manufacturing

Encodes maintenance logs and SOPs; frontline staff query in plain language.

Healthcare

Summarizes clinical protocols under HIPAA-grade on-premise controls.

Mid-market ops

Onboards new hires against your real playbook instead of generic training.

R&D & engineering

Searches proprietary design history and internal specs in seconds.

The comparison

What actually changes when you own the model

A public API and a private model can answer the same question. They are not the same purchase. Here is what's actually different underneath.

Dimension Public API 1xNEO
Where data is processed Third-party cloud servers Hardware inside your building
What it knows General internet, plus your prompts Fine-tuned on your own documents
Pricing model Per-token, scales with volume Owned hardware, flat going forward
Cost as usage grows Rises linearly, no ceiling Improves; the hardware is already paid for
Works without internet No, requires a live connection Yes, air-gap capable
Who can see prompts and outputs Governed by the vendor's policies Only the team you scope access to
Model changes Vendor-controlled, can change anytime Retrained on your schedule, on your data
What you end up with Continued access, for as long as you pay Hardware and model weights you own outright

Delivery

Four phases from first conversation to daily use

01

Discovery & data audit

Two days on site mapping every document store, ticket queue and spreadsheet that carries real institutional knowledge. We agree on what stays out.

02

Model training

An open-weight base model is fine-tuned on your corpus inside an isolated environment. Evaluation sets are written by your domain experts.

03

On-premise install

We deliver and rack the inference hardware in your office, wire it to your network, and hand over the admin console. Internet access is optional after this.

04

Handover & ongoing care

Staff trained on prompting and review workflows. Quarterly retraining and SLA-backed support keep the model sharp as you grow.

"Our compliance officer signed off in one meeting. The model reads twelve years of case files and the answer never leaves the third floor. It performs like the public tools we'd been told we couldn't use, without the risk."

Operations lead · 60-person legal practice, Copenhagen

FAQ

Questions worth asking before you commit

How is this different from ChatGPT Enterprise or Azure OpenAI?

Those are still multi-tenant cloud services. Your prompts and documents travel to a third party's servers no matter how the contract is worded. 1xNEO runs on hardware physically inside your building. There is no tenancy to audit and no data-processing agreement to trust, because there is no data leaving in the first place.

What models do you actually use?

We fine-tune open-weight foundation models. Llama, Qwen, Mistral and DeepSeek are common choices, depending on your workload and hardware footprint. You are never locked into one lab's roadmap; if a better open-weight model ships, we can retrain on it.

How does it compare to GPT-4 or Claude on quality?

For narrow, company-specific tasks (summarizing your contracts, answering from your internal wiki, drafting in your house style) a fine-tuned model matches or beats a general-purpose frontier model, because it isn't guessing from the general internet. For open-ended reasoning across unfamiliar domains, frontier models still lead. Most clients keep a frontier API for that slice of work and run everything else privately.

Are we locked into 1xNEO once we deploy?

No. The hardware is yours outright, the model weights are yours, and the fine-tuning pipeline is documented and handed over. If you ever want to run it without us, you can.

Is this a subscription, a one-time purchase, or both?

The hardware is a one-time purchase you own outright. Keeping the model current (quarterly retraining, support and monitoring) is an ongoing arrangement. You are never billed per token or per query.

What if the hardware fails?

Deployments include a maintenance plan with on-site service and parts. Larger deployments run a two-node cluster with automatic failover, so a single hardware fault doesn't take the model offline.

Can it run fully offline?

Yes. Air-gapped deployments are supported. Model updates are delivered on encrypted media instead of over the internet, for teams that cannot allow any outbound connection at all.

How long does deployment take?

Roughly six weeks from the first data audit to staff actively using the model day to day: two days on-site for discovery, a few weeks of training and evaluation, then install and handover.

Who can access the model once it's deployed?

Access is role-scoped from day one. Only the teams you designate can query it, every request is logged, and permissions look and feel like any other internal system your IT team already manages.

Book a demo

Tell us what your team keeps re-explaining

Send a short brief and we'll come back with a feasibility read, an indicative hardware spec and a timeline. Scoping calls run 45 minutes and cost nothing.

hello@1xneo.com

Mon – Fri · 09:00–18:00 CET

Opens your email client with this brief pre-filled.