REX42
Try it live Talk to us
Deployment

Small models run where large models cannot.

A model that requires a rack of accelerators cannot run in a clinic, a warehouse office, a shipping agent’s back room, or a government department that is not permitted to send data outside the building. REX42 runs on the hardware those places already own.

2 GB+
Any VPS or cloud server
CPU-only
No GPU required
~200 MB
Storage at 4-bit
Air-gapped
Supported
wide image
Three ways to run it

Same weights, same behaviour — only the location changes.

Mode 01
On your own server
Inside your network, behind your firewall, managed by your IT team. The most common choice for document workloads where the archive already lives on-premises.
Mode 02
In your private cloud
Your tenancy, your region, your keys. Useful when you have no on-premises hardware but still cannot send data to a third-party AI provider.
Mode 03
On the device itself
A laptop for a field team, an edge box on a factory floor, or a handset. Inference happens where the data is captured and nothing needs to leave.

In all three cases there is no outbound call path at inference time. The deployment mode is an operational decision, not a privacy trade-off.

Devices it can run on

Tested, with honest expectations for each.

Where a device needs 4-bit quantisation to fit, the table says so. Where it will not work at all, the table says that too.

Device Runs it? Expected speed
Any VPS / cloud server (2 GB+) Yes Good
Mini PC (Intel NUC, N100 box) Yes Good
Laptop / desktop, any modern CPU Yes Fast
Raspberry Pi 5 (4 or 8 GB) Yes, 4-bit Usable — a short answer in a few seconds
Raspberry Pi 4 (4 GB) Yes, 4-bit Slow but works
NVIDIA Jetson Orin Nano Yes, GPU Fast
Android phone, 6 GB+ RAM Yes, 4-bit Usable
iPhone 12 and newer Yes, 4-bit Usable
In the browser (WebAssembly / WebGPU) Yes ~200 MB download first
Microcontrollers (ESP32, Arduino) No Nowhere near
A Raspberry Pi 5 answering a document question in a few seconds is the clearest statement of what small means. Nothing in this table is a GPU cluster.
Where it runs

Three deployments, one set of weights

n image
On-prem
On your own server
Inside your network, behind your firewall, managed by your IT team.
n image
Private cloud
In your tenancy
Your region, your keys, no third-party AI provider in the path.
n image
On-device
On the device itself
A laptop, an edge box, or a handset. Inference where the data is captured.
Hosting it for a website

The model runs on your server, not the visitor’s device.

If you are putting REX42 behind a public-facing product, these are the sizing bands we use.

Users Server Cost
Demo / light use 2 vCPU, 4 GB RAM, no GPU ~$20 / month
Steady business use 4–8 vCPU, 8 GB RAM ~$40–80 / month
Heavy or fast replies 1 × T4 or L4 GPU ~$0.35–0.50 / hour
Compare that to a per-user, per-month subscription for a hosted frontier model. The saving is structural, and it lands with you.

For in-device deployments this table does not apply at all — there is no server, and the marginal cost of an additional user is zero.

How it connects

It behaves like any other internal service.

REX42 exposes a local REST and queue interface. Your ERP, CRM, DMS or file share talks to it the same way it talks to anything else on the network.

POST /v1/extract
{
  "task": "invoice",
  "file": "base64:JVBERi0xLjQK...",
  "fields": ["supplier", "total", "vat", "date"]
}
200 OK — application/json
{
  "supplier": "Gulf Marine Services LLC",
  "total": 18450.00,
  "vat": 878.57,
  "date": "2026-08-14",
  "confidence": 0.94,
  "unreadable": []
}
Structured output only. Every task returns JSON against a schema you define. No prose to parse, no prompt engineering in your application code.
Confidence is returned, not hidden. Low-confidence values are flagged rather than guessed, and you set the threshold at which they route to a human queue.
Local, structured logging. Every inference writes input hash, model version, latency and output fields to a log that stays in your infrastructure and plugs into your existing SIEM.
No SDK lock-in. It is an HTTP endpoint on your own network. Any language that can make a request can use it.
Integration

It behaves like any other internal service.

A local REST and queue interface on your own network, returning JSON against a schema you define. Your ERP, CRM, DMS or file share talks to it exactly the way it talks to anything else.

n image
What deployment actually looks like

Two to six weeks, depending on data readiness.

STEP 1
Sizing
We measure your real volumes and formats and tell you what hardware is genuinely needed — usually less than expected.
STEP 2
Install
Model, runtime and weights install from local media. For air-gapped sites we commission on-site with your IT team.
STEP 3
Integration
The endpoint is wired into the system that will consume it, with logging pointed at your monitoring.
STEP 4
Handover
You get the weights, the licence, a signed install manifest and a runbook. Support afterwards is optional, not structural.

You can run it yourself from day one. The licence is perpetual and the system keeps working if the contract ends — that is a deliberate property of the architecture, not a commercial gesture.

Keep reading

Tell us the hardware you have.

We will size the deployment against your real volumes before you buy anything.

Talk to us Try it live