REX42
Try it live Talk to us
REX42 private inference unit
Private AI · Built in Dubai

Custom AI that does your job,
runs on your devices, and
keeps your data with you.

REX42 builds small, private language models that do specific jobs for your company — on your server, in your private cloud, or on the device itself. No data leaves your premises. No per-user fees. No dependency on ChatGPT, Claude or Gemini.

Try the Playground Talk to us about a custom model
350M params CPU only No data egress
350M
Parameters
CPU only
No GPU required
On-prem · Private cloud · On-device
Runs where you need it
Perpetual licence
Your own deployment
Vault holding private documents and records
Private by design

Every business wants AI.
Very few can afford to send
their data to it.

Your documents, customer records and internal data belong inside your business.
REX42 brings the AI to them.

REX42 — the third option.
Why private AI matters
What it is

REX42 is not
a chatbot.

A worker trained for your job. You pass data in through an API and it replies in JSON — no conversation to manage, no prompt engineering in your application code.

One task, one answer. Trained for your formats rather than adapted from something general.
Structured output only. JSON against a schema you define, ready for your ERP, CRM or DMS.
Honest when it cannot read. Low-confidence values are flagged, never invented.
See what it can do 15 live capabilities
Input
Your document
Contracts, invoices, IDs, CVs or records — a file, a photograph or raw text.
POST /v1/extract
The model
REX42, on your hardware
Extracts · classifies · corrects · triggers. 366M parameters, CPU only, no outbound call.
200 OK · application/json
Output
Structured data
Named fields, a confidence value, and an explicit list of anything it could not read.
 
{
  "supplier": "Gulf Marine Services LLC",
  "total": 18450.00,
  "vat": 878.57,
  "confidence": 0.94,
  "unreadable": []
}
Capabilities

Small model. Real work.

Fifteen tasks are live today, each one trained for a specific job rather than prompted out of a general model. Every task returns structured JSON, and every one runs on your own hardware.

Explore all capabilities
Text processing
Summarise, rewrite and understand your content.
Summarise and rewrite
Grammar check and tone
Sentiment and intent routing
Document data extraction
Turn IDs, invoices and contracts into structured fields.
Emirates ID, passport, visiting card
Invoices and receipts
CVs and tenancy contracts
Data intelligence
Find duplicates. Search by meaning.
Duplicate customer detection
Meeting notes into actions
Accounting head classification
Deployment

Small models run where large models cannot.

Explore deployment
Tower server with red light strip
On-prem
On your own server
Windows or Linux, CPU only — no accelerator budget and no procurement cycle. It sits behind your firewall and your IT team manages it like any other internal service.
Rack cluster of private cloud servers
Private cloud
In your private cloud
Your tenancy, your region, your keys. The right answer when you have no on-premises hardware but still cannot hand documents to a third-party AI provider.
Edge boxes and a phone running REX42 offline REX42 running on the device
On-device
On the device itself
A laptop, an edge box or a handset — about 200 MB of storage at 4-bit. Inference happens where the data is captured, with the network cable out.
Architecture

Many small models.
One low-cost system.

Specialist models, each trained for one purpose. Together, they cover your workflow on your own hardware.

How it works
Your server · CPU only
Invoice reader
Customer messages
Semantic search
ID extraction
Duplicate detection
Summariser
In production

It runs inside Royex's own
Agentic Business Suite today.

REX42 is not a lab experiment. It is already the private-inference layer inside Royex's Agentic Business Suite, reading the documents our own clients would never send to a hosted model — while public models keep handling the work where privacy is not the constraint.

The router decides per task, not per product. Anything carrying identity documents, contracts, salaries or customer records stays on the local model; general drafting and open-ended questions go out to a large model. One workflow, two engines, and a clear line between them.

How the line is drawn Running in production since 2025
Sensitive tasks → REX42
Emirates IDs, passports, invoices, contracts, CVs, payroll and customer records. Read on our own hardware, never transmitted, with every field checked against the source document.
General tasks → Public models
Marketing copy, research, brainstorming and general drafting — work where nothing confidential is involved and a frontier model is simply the better tool.
Get started

Ready for AI that works for your business?

Tell us the task. We will tell you how small the model can be — and where it can run.

Book a discovery call Open the Playground
FAQ

Frequently asked questions.

The questions security, IT and finance teams ask us first. Anything else, our engineers will answer directly.

Ask us a question

No. The model runs on hardware you control — your server, your private cloud tenancy, or the device itself. Nothing is sent to an external API.

No. REX42 is a 366M-parameter model tuned for CPU inference on ordinary Windows or Linux servers.

A perpetual licence for your own deployment. No per-user or per-token fees.