Hardware and model computing capacity

Company-trained local AI models.
Optimised hardware stations.

Core 1Models & stations Core 2Agentic OS Core 3Tailored automation

Frontier-class answers for a whole team, on one machine, paid once. No model is called over the internet. No usage is metered.

The 3 tiers

How many people one station serves at once.

Every station carries all three tiers. You choose per task. The numbers are the capacity for daily agentic work.

Ultra-fast

For quick work: sorting, classifying, extracting, quick answers. Faster than you can read.

At once: up to 6 people per station

Medium

For everyday work: mail, documents, customers, multi-step agent tasks. It runs most of the day.

At once: up to 3 people per station

Smart

For the hardest cases: whole case files and decisions you cannot get wrong.

At once: 1 person per station

A station runs one tier at a time. More people means more stations, pooled.

Day and night

The same capacity, twice used.

A station that serves people by day is not idle at night. One purchase does two jobs.

Working hours · agentic use

People work in the apps and ask the agent in plain language. It prepares. They approve.

Outside working hours · background automation

Routines on a schedule. Document processing, onboarding queues, screening, reconciliation, morning reports.

Control

Every background job is a routine you can see, pause and audit. Results wait for a person.

Sealed either way

Day or night, every AI action is written to the Trust Log first. The protocol →

The machines

3 sizes of the same system.

Same Clerk OS, same apps, same three tiers. What changes is how many people and how much data at once.

For an office or a team

Clerk Station

The out-of-the-box machine. 6 people at once on ultra-fast, 3 on medium, 1 on smart.

Graphics: 2 × RTX 5070 Ti · 32 GB
Memory: 64 GB · 96 GB for AI
Storage: 1 TB NVMe
Running cost: about €63 a month

from €5,210paid once · founding price · list €6,130 · editions from €6,039

The stations, specification and price →

For a department or a larger firm

Clerk Machine

The same Clerk OS with 3 times the memory and 2 times the graphics power.

Graphics: 2 × RTX 5090 · 64 GB
Memory: 192 GB DDR5
Storage: 3 TB SSD

Price on requestquoted in writing

For an institution

Clerk Custom Node

Built for one organisation. Single nodes, clusters or air-gapped, with your own model on it.

Sizing: To your people and load
Network: Air-gapped if required
Model: Fine-tuned on your data

Scoped in writinga pilot first

Cost shape	Hosted copilots	Clerk Station
Per seat	About USD 30 per seat per month, on top of USD 39–60 for the office suite (vendor list prices, 2026).	No per-seat fee. Paid once. About €63 a month of electricity.
Grows with	Headcount, every month.	Machines you add. 6 people at once per station on ultra-fast.
Who holds the switch	The vendor: models retired, caps and prices changed, accounts suspended.	You. No account, no licence server, no renewal.

Every kind of model a business uses, installed

Language & reasoningEvery station

VisionEvery station

OCR & document readingEvery station

TranscriptionEvery station

Voice generationEvery station

Image generationEvery station

Search over your documentsEvery station

Video generationCreator edition

Music & soundCreator edition

3D modelsCreator edition

UpdatesA better open model is tuned for Clerk OS and installed as an update, offline if you like.

Included or optionalMaintenance updates are included. Packs with the next model generation are optional.

Fine-tuning on your data

A model you own, trained on how your institution works.

For high-precision work, a general model is not enough. We build you one that performs best at your specific work.

What goes inDocuments, filings, correspondence, templates, the rules you apply, past decisions. Only what you clear is used.

What it learnsYour terminology, your document structures, how your reviewers decide.

What you receiveThe model file, the training record, results on your held-out documents. Agreed as acceptance criteria.

Use case for high precision	What the fine-tuned model does better
Supervisory filings	Reads each filing against the rules on arrival. Writes the findings in the house wording.
Credit and claims decisions	Assembles the decision memo the way your reviewers write it.
Classification	Document types, ledger entries, case categories, in your own taxonomy.
Screening decisions	Applies your thresholds and escalation rules, with the evidence laid out.
Correspondence	Letters, replies and reports in your house style, in every language you work in.

On your premisesTraining runs on the nodes we install, inside your network. Nothing is uploaded.

Retrained under supportWhen the rules, the forms or the business change.

Yours to keepLeaving means keeping the model, the integrations and the audit trail.

Talk to us about your model →Scoped in writing · a pilot on a first cleared set of documents

Specification

The 3 tiers, measured.

On the station, October 2026, through the product's own code path. Context is 200K+ tokens on every tier.

Tier	Model class	Writing	Reading	At once
Ultra-fast	Mixture of experts · about 35B, 3B active per token	about 220 tokens/s	up to 11,000 tokens/s	up to 6
Medium	Dense · about 27B parameters	45–130 tokens/s	—	up to 3
Smart	Dense · 200B-class	100+ tokens/s	about 5,000 tokens/s	1

Ultra-fast reads a 20-page brief in about a second. Medium trails the largest rented models only on world knowledge.

Core 1 of 3

Own your intelligence.

The models, the machines and the switch, in your building. Next: the agentic OS that runs on them.

Core 2 · the out-of-the-box apps →Core 3 · the methodology →

info@finqup.com · a person answers

Company-trained local AI models.Optimised hardware stations.