all models local Buy

Now taking preorders


Build yours How it compares

£250 deposit · fully refundable · nothing charged until it ships

0x7F3A9C21
  • NET switch: Cuts the network physically. Flip it and nothing can leave.
  • 4 × Thunderbolt: Model storage, displays and capture, all connected at once.
  • 2 × 10GbE: Serves the whole office over your own network.
  • Machined fin block: Solid aluminium plates carry heat away under sustained load.
  • Touch display: OLED screen, wakes when touched: model, speed, memory, temperature, and bytes sent. Always 0.
  • Stacking mounts: The feet of one unit seat into the top of the next. No cable.

A low, wide satin-aluminium AI computer with a machined fin block at one end, four stacking feet, and a rear panel of Thunderbolt, 10GbE, USB, HDMI and DisplayPort ports beside a physical network isolation switch.

  • No cloud fallback
  • Physical network switch
  • Models pre-installed
  • Up to 128 GB unified memory
  • Installation included
  • Stacks with no cable

The situation

Three ways to get AI.
One stays home.

Most businesses either send private material to a service they do not control, or skip AI altogether. The "local" boxes in between often quietly route the hard questions to the cloud.

Cloud AI

Reads everything you give it

Every document, client name and draft contract goes to someone else's servers, under terms they can change. Plenty of firms are not allowed to do that. Not reluctant: not allowed.

Hybrid "local" boxes

Local until it isn't

The agent lives on a box in your office, but the thinking is routed to cloud models whenever a task is too big for it. For law, health and finance, "sometimes it goes to the cloud" fails the compliance question outright.

Do it yourself

A weekend of Docker errors

The open tools are good, and built for engineers: drivers, quantisation formats, inference servers. Most people give up before anything works.

cato8

All of it on the machine. A switch that proves it.

Up to 128 GB of unified memory holds models to 120B on the device itself, so there is never a reason to send anything out. Flip the NET switch on the back and keep working. That is the only demonstration of privacy that proves anything.

See the full comparison

0 GB

unified memory for models

0 B

parameters, run locally

0

outbound requests, by design

0 s

from first boot to first answer

What you're actually buying

The hardware is ordinary.
cato8 OS is not.

Anyone can put fast silicon in a small box. The operating system is the product, built for running models, not for running a desktop.

cato8 OS Offline
22.4 tokens / sec
61GB memory in use
0 outbound requests

Models arrive installed

Weights are on the drive when it ships. First boot to first answer is about ninety seconds: no download, no account.

One place for everything

Swap models, add a pack, connect a folder. No terminal, no config files.

Built to be left alone

Signed updates and an immutable system partition. An appliance that behaves like one.

Nothing phones home

No licence check, no telemetry on what you ask, no account required. Turn the network off and it carries on.

In your pocket

Run it from your phone.
Not through our servers.

The cato8 app talks to your machine directly, over your own network or your own VPN. Other AI boxes route your phone through the maker's relay. Ours has nothing in the middle to route through.

  • Chat and answers Ask from anywhere in the building. Answers cite the documents they came from.
  • Approve before it acts Agents ask before they send, delete or change anything. One tap to approve.
  • Agents that keep working Scheduled jobs and long tasks run on the machine when your laptop is closed.
  • Every action logged What ran, when, on which model and which documents. Exportable.

Integrations

Works with what you already use.

Your email, files, calendar and business systems, connected so the machine can read and act on them. The thinking still happens on the machine; the logins never leave it.

M3 Microsoft 365 O Outlook T Teams S SharePoint O OneDrive GW Google Workspace G Gmail GD Google Drive S Slack N Notion D Dropbox B Box
S Salesforce H HubSpot J Jira C Confluence G GitHub G GitLab X Xero Q QuickBooks C Clio SD Shared drives (SMB) IA IMAP and CalDAV P PostgreSQL AM Any MCP server

Connected for you, on-site

A SetupMyAI engineer connects the systems you actually run, then tests them with your team.

Logins stay on the machine

Credentials are stored on your hardware and never pass through the model. Revoke any of them in one place.

Open standards

Connections use MCP and each system’s own API. If it has an API, it can usually be connected.

Your choice of silicon

NVIDIA for business.
AMD for personal.

Business tiers run on NVIDIA Grace Blackwell, for CUDA and the fastest work on long documents. Personal runs on AMD to keep it affordable, and can be specified with NVIDIA. The operating system, models, NET switch and engineer are the same on both.

Standard in Forum, Curia and Senate

GB10 Grace Blackwell

The superchip behind DGX Spark. Full CUDA, and about five times faster than AMD at reading long documents.

  • 128 GB unified memory
  • 20-core Arm CPU, Blackwell GPU
  • Up to 1 petaFLOP FP4, NVIDIA figure

Standard in Atrium · lower-cost business option

Ryzen AI Max+ 395

Unified memory shared by CPU and GPU, so a large model loads whole. The most model per pound.

  • 64 or 128 GB unified memory
  • 16 Zen 5 cores, Radeon 8060S
  • Same models, slower on long prompts

Enterprise · built to order

RTX PRO Blackwell

Workstation GPUs for the fastest generation on models that fit in graphics memory, and multi-GPU racks.

  • Up to 96 GB GDDR7 per card
  • Multi-GPU workstations and racks
  • Priced on request

Chip specifications are the manufacturers' published figures. Every option and its price is on the configurator.

Data sovereignty

Unplug it.
It still works.

For a lot of our customers this isn't a privacy preference. Their material is not permitted to leave the building, privileged case files, patient records, unpublished results, work under NDA, anything inside a regulator's perimeter. "The provider says they don't train on it" is not an answer to that. Not sending it at all is.

The models live on the device, so there's no inference call to make and nothing to send. Pull the network cable out of the back and keep working, which is the only demonstration of privacy that actually proves anything.

  • No cloud inference, not as a fallback, not for "improvements"
  • A physical switch on the back that cuts the network, provably
  • A display on the front that counts bytes sent. It reads 0
  • No account required to use the machine
  • Deployable inside genuinely air-gapped environments
  • Takes the AI processor out of your GDPR position entirely
The rear panel in satin black: DC in, two 10GbE ports, four Thunderbolt, USB, HDMI, DisplayPort and the NET isolation switch, each labelled
Network Disconnected
AssistantRunning
Document searchRunning
VoiceRunning

Outbound requests since boot, 0

Your own local intelligence

It learns your work.
It keeps it to itself.

cato8 reads your documents, email and notes where they live and keeps its index on its own drive. The more of your work it holds, the more useful its answers get, and every answer shows the document it came from.

Compatible with your devices

Anything on your network can use it, from a browser. Displays, storage and microphones plug straight into the back.

  • Mac and Windows
  • Linux
  • iPhone and Android
  • iPad and tablets
  • Thunderbolt displays
  • External model storage
  • Microphones, for voice
  • Your office network

Find your setup

Three questions. One answer.

Tell us the job and we'll work out the machine, the models and the setup. No form, no email, no sales call.

01 / 03

What do you want it to do?

Choose as many as apply.

Packages

One for you. Three for business.

Compare all specifications
cato8 Atrium package: AMD Ryzen AI Max+ 395, satin finish, machined fin block

Personal · for one person

Atrium

A strong private assistant that runs easily, all day, on your desk. NVIDIA available as an upgrade.

Largest model 32B
8B70B671B

£1,500 from, excl. VAT, installation included

  • Chip AMD Ryzen AI Max+ 395
  • Memory 64 GB unified
  • Runs Qwen3 32B, Llama 3.1 8B
  • People One
Configure
Most chosen
cato8 Forum package: NVIDIA GB10 Grace Blackwell, satin finish, machined fin block

Business · a small team

Forum

One NVIDIA machine for a small team: 70B and 120B models, company knowledge, your systems connected.

Largest model 120B
8B70B671B

£4,199 from, excl. VAT, installation included

  • Chip NVIDIA GB10 Grace Blackwell
  • Memory 128 GB unified
  • Runs gpt-oss-120b, Llama 3.3 70B
  • People About five
Configure
cato8 Curia package: 2 × NVIDIA GB10, satin finish, machined fin block

Business · an office

Curia

Two NVIDIA machines stacked into one 256 GB pool, or run side by side for twice the people.

Largest model 235B
8B70B671B

£8,398 from, excl. VAT, installation included

  • Chip 2 × NVIDIA GB10
  • Memory 256 GB pooled
  • Runs Qwen3 235B pooled
  • People About ten as two nodes
Configure
cato8 Senate package: 4 × NVIDIA GB10, satin finish, machined fin block

Business · a department

Senate

Four NVIDIA machines and 512 GB: the largest open models, or a whole department at full speed.

Largest model 671B
8B70B671B

£16,796 from, excl. VAT, installation included

  • Chip 4 × NVIDIA GB10
  • Memory 512 GB pooled
  • Runs DeepSeek-R1 671B pooled
  • People About twenty as four nodes
Configure

For an organisation

Enterprise. Custom quote.

Frontier-size open models across pooled machines or NVIDIA Blackwell builds, per-user access control, exportable audit logs, air-gapped deployment and a named engineer. Scoped with you, then quoted.

Request a quote Typically a 2–6 week deployment

Prices exclude VAT and include installation by a SetupMyAI engineer, on-site across the UK and UAE, remotely everywhere else. UK orders have 20% VAT added at checkout; orders shipped outside the UK are zero-rated and pay the price shown. Business tiers are priced per machine and run on NVIDIA as standard, with AMD as a lower-cost option. You are not buying a box; you are buying a working deployment.

Tech specs

What is inside, and what it runs.

Every package in its standard build, component by component, with the open models it runs and roughly how fast.

cato8 Forum: one machine
Chip
NVIDIA GB10 Grace Blackwell
CPU
20 Arm cores per machine (10 Cortex-X925, 10 Cortex-A725)
GPU
Blackwell, up to 1 petaFLOP FP4 per machine · CUDA
Memory
128 GB LPDDR5X unified, shared by CPU and GPU
Bandwidth
273 GB/s per machine
Storage
2 TB SSD per machine, up to 8 TB
Wireless
None as standard · Wi-Fi 7 and Bluetooth optional, cut by the NET switch
Rear I/O
4 × Thunderbolt · 2 × 10GbE · 2 × USB-A · HDMI · DisplayPort · DC in
Isolation
Physical NET switch · Kensington slot
Front display
OLED screen, wakes on touch. Silver mirror on aluminium, black glass on black
Stacking
Feet and top receptacles with gold contacts
People
About five

Models it runs, 4-bit

  • gpt-oss-120b ~40 tok/s
  • Llama 3.3 70B ~5 tok/s
  • Qwen3 32B ~11 tok/s
  • Whisper Large v3 voice

Chip figures are NVIDIA's and AMD's published specifications. Speeds are approximate generation rates from the model table below; NVIDIA builds generate at a similar rate and read long prompts about five times faster. Pre-launch: we will publish measured numbers on shipping units. Every option is on the configurator.

The machine, up close

Machined, down to the contacts.

Satin aluminium or satin black, a fin block cut from solid, every rear port labelled, and gold contacts where one machine stacks onto the next.

The front display awake, showing Llama 3.3 70B, 5.1 tokens per second, memory, temperature and 0 bytes sent
Touch to wake An OLED screen on the front: model, speed, memory, temperature, and bytes sent. Always 0. Silver mirror on aluminium.
cato8 in satin aluminium, front three-quarter view
Satin aluminium Brushed vertically, with a single crisp chamfer on every edge.
Close-up of the rear ports and the NET isolation switch
Every port, labelled Four Thunderbolt, two 10GbE, USB, HDMI, DisplayPort and the NET switch.
The underside of a cato8, showing four feet with gold contacts
Gold in every foot The feet are the stacking connector. No cable, no switch.
The top of a cato8, with four stacking receptacles
Receptacles on top Machined sockets on the same centres as the feet, for the unit above.
cato8 in satin black, front three-quarter view
Satin black The same machine in near-black, its form carried by the highlights.
The front face of a cato8 with its wordmark and status light
A clean front Wordmark, node code and one status light. Nothing else.

Outgrown it

Need more? Add another one.

Every machine can be linked to another. Two units pool their memory into one larger space, so a model that fits on neither will run across both. You are never buying the ceiling on day one, and you are never throwing the first machine away to get past it.

Two cato8 machines stacked, the upper unit's feet seated in the receptacles of the one below
What it does

Bigger models become possible

Memory is the hard wall in local AI. Pool two Professional machines and you have 256 GB, which opens up models that simply will not load on 128 GB. This is the main reason to stack.

What it does

More people at full speed

Run them as two independent nodes instead and each serves its own users at full speed. For a growing team this is the genuine doubling, and it is usually what people actually need.

What it does not do

It will not make one answer twice as fast

We would rather say this up front. When a single model is split across two machines, the link between them carries a fraction of what internal memory does, so the pair generates at roughly the speed of one. You gain capacity and headroom, not raw speed on a single query. Anyone promising you double throughput is selling.

Model compatibility

Exactly what each configuration runs.

Named models, 4-bit, with approximate generation speed. No asterisks: if it says it runs, it runs, and where it will not fit we say so.

Model Personal64 GB Professional128 GB Two stacked256 GB Enterprise512 GB+
Llama 3.1 8B8B · ~5 GB ~35 tok/s~40 tok/s~40 tok/s~40 tok/s
Qwen3 32B32B · ~19 GB ~10 tok/s~11 tok/s~11 tok/s~12 tok/s
Llama 3.3 70B70B · ~40 GB Will not fit~5 tok/s~5 tok/s~6 tok/s
gpt-oss-120b117B mixture of experts · ~65 GB Will not fit~40 tok/s~40 tok/s~45 tok/s
Qwen3 235B235B mixture of experts · ~120 GB Will not fitVery tight~10 tok/s~14 tok/s
DeepSeek-R1 671B671B mixture of experts · 350 GB+ Will not fitWill not fitNeeds four~12 tok/s

Mixture-of-experts models are large on disk but activate only a fraction of themselves per token, which is why the 120B runs so much faster than the 70B. Long context costs memory on top: budget roughly a quarter of the model again for everyday work, and closer to double if you intend to work at 128k tokens.

Llama 3.3 70B gpt-oss-120b Qwen3 235B Qwen3 32B Qwen 2.5 Coder 32B Mistral Small 24B Llama 3.1 8B DeepSeek-R1 671B Whisper Large v3

AI Packs

Choose what it's for. It ships that way.

A pack is a working configuration, models, applications and settings chosen for a job, installed before the machine leaves us.

Personal

Assistant, private documents, voice, knowledge base.

Developer

Coding models, local APIs, model management.

Creator

Image generation, audio, video, creative tools.

Business

Company knowledge, document intelligence, automation.

Enterprise

Multi-user, access control, monitoring, integrations.

SetupMyAI

You don't have to do any of this yourself.

SetupMyAI is our deployment arm. Engineers come to your office, install the machine, connect it to your systems, train your team and stay until it works, on-site across the UK and UAE, remotely anywhere else in the world.

Get your AI set up For business
  • Personal Delivered, installed, connected to your files, explained properly.
  • Business Company knowledge base, integration with your systems, staff training.
  • Enterprise On-premise deployment, multi-node, access control, ongoing support.

Where this goes

Every machine is a node.

Compute today is concentrated in a handful of data centres. We think it should be distributed across the people who own it. Machines already cooperate on a local network; the protocol that lets them cooperate more widely is being built in the open.

Available today

A single machine, yours, running locally.

In development

Several machines in one office, pooled into one larger model.

Research

Federated compute between independently owned nodes.

We're not selling this. It is not a return on your hardware. It's the direction we're building in, and we'd rather say so plainly.

Questions

Asked, and answered plainly.

Anything else, ask us. A person reads every email.

Does cato8 ever send my data to the cloud?

No. The models run on the machine itself, so there is no cloud inference to fall back to, not even for large models. A physical NET switch on the back disconnects the network entirely, and the machine keeps working with the cable unplugged. No account is needed to use it.

What AI models can cato8 run?

Atrium (64 GB) runs 30B-class models such as Qwen3 32B comfortably. Forum (128 GB) runs Llama 3.3 70B and the 120B mixture-of-experts gpt-oss-120b. Curia pools 256 GB and runs Qwen3 235B; Senate pools 512 GB and runs DeepSeek-R1 671B. Models come pre-installed on the drive.

How much does cato8 cost?

Atrium, for one person, starts at £1,500 on AMD. The business machines run on NVIDIA GB10 Grace Blackwell at £4,199 per machine: Forum (one machine, £4,199), Curia (two, £8,398) and Senate (four, £16,796), with an AMD option that costs less. Memory, storage, wireless and cover can all be configured, Apple-style, at /buy. Prices exclude VAT and include installation by a SetupMyAI engineer. UK orders add 20% VAT; orders delivered outside the UK are zero-rated. Enterprise is a custom quote.

Is there a subscription?

No. There is no subscription, activation charge or setup fee needed to use the machine. Running it harder costs nothing beyond electricity.

How does the preorder work?

A £250 deposit reserves your place. It is fully refundable at any time before your machine ships, for any reason. Your price is fixed when you preorder, and nothing else is charged until the machine is ready to ship. We have not committed to a shipping date and will tell you directly when we have one we believe in.

Can cato8 use NVIDIA chips?

Yes. Forum, Curia and Senate run on NVIDIA GB10 Grace Blackwell as standard: 128 GB of unified memory per machine, full CUDA, and much faster on long documents. Atrium runs on AMD Ryzen AI Max+ to keep it affordable, and can be specified with NVIDIA. RTX PRO Blackwell builds are available for Enterprise. The operating system, pre-installed models, NET switch and installation are the same on every build.

Who installs it?

A SetupMyAI engineer, included in the price. Installation is on-site across the UK and UAE and remote everywhere else. They connect it to your files and systems and train your team.

Does stacking two machines make it twice as fast?

No. Stacking pools memory, so two units can run models that neither could load alone, and two units run as separate nodes can serve twice as many people at full speed. A single answer from one model split across two machines generates at roughly the speed of one unit.

What happens to my machine if cato8 stops trading?

It keeps working. Models are on the drive, the weights are open, the operating system does not check a licence and nothing phones home. If we cancel the product or stop trading before your machine ships, deposits are refunded.

What warranty does it have?

A one year limited warranty against defects in materials and workmanship, from delivery. It is in addition to your statutory rights, which in the UK run for up to six years under the Consumer Rights Act 2015 and are not limited by the warranty.

Stop renting
your intelligence.

A chat subscription costs tens of pounds a month. Running agents that read documents all day, write code and answer from your own files is metered on tokens, and the bill lands in the thousands.

10 people

Cloud, per month

  • Seat licences£450–£900
  • Agent & API usage£800–£4,000

£1,250–£4,900

every month, indefinitely

Owned, once

  • 2 × Professional, excl. VAT£5,798
  • Usage thereafter£0

£0

every month, thereafter

Covered in 2–5 months at these rates.

Assumes £45–£90 per seat, £80–£400 of agent usage per person, and one Professional per five people.

Find your setup Talk to us

Illustrative, and it depends entirely on what you run. These figures assume a small team running agents rather than occasionally chatting: at published frontier-model API rates, one always-on document workflow costs several hundred pounds a month, and a coding assistant driven all day by a single developer can exceed that on its own. If your usage really is a few questions a day, the cloud is cheaper and we'll tell you so.