On-Prem

Transkribus, deployed on your infrastructure.

Text recognition, layout analysis, table extraction, NER, and model training — deployed on your servers, tailored to your documents.

Text RecognitionLayout AnalysisTables & FieldsNERModel TrainingAir-Gapped

transkribus-onprem — recognition

$ transkribus recognize --model super-v2 ./docs/Loading model super-v2 …GPU: NVIDIA RTX 4090 (24 GB) — CUDA 12.4Processing 847 pages├─ Layout analysis ····· done (0.8s/page)├─ Text recognition ··· done (3.1s/page)├─ Table extraction ··· done (0.4s/page)├─ NER ················ done (0.2s/page)└─ Export: PageXML, PDF✓ 847 pages — CER 2.1% — 43 min Results: ./output/pagexml/

What's included

The on-prem package gives your institution the full Transkribus platform, running entirely on your own hardware.

Recognition — Super Model v2

Den 15ten Januar 1847

wurde in der Kirche zu

Weimar das Kind Johann

Friedrich getauft...

geboren am 12ten des

Monats, Sohn des...

CER2.1%

Text Recognition

Handwritten and printed text recognition with Standard HTR and Super Models. The same engines used on the Transkribus cloud — identical accuracy, running on your GPU.

Standard HTR for high-throughput batch processing
Super Models for 70+ scripts and challenging material
Per-character confidence scores and CER reporting
GPU-accelerated — ~3 seconds/page on RTX 3090

Complete data sovereignty

Your documents never leave your network. No cloud dependency, no external data transfer, no third-party access — by design.

Air-gapped operation

After initial setup, no internet connection is required. The entire platform runs locally — recognition, training, and export all happen on your hardware.

Your infrastructure, your rules

You control storage locations, retention policies, access permissions, and backup schedules. Transkribus runs as containers on your servers — nothing is shared.

Controlled updates

Model and platform updates are delivered as container images. You decide when and whether to update. No forced upgrades, no surprise changes.

Compliance ready

Designed for institutions with strict data handling requirements. No data leaves your premises — suitable for classified, sensitive, or legally restricted material.

Your premises

Your infrastructure

Upload

↓

Recognition on your GPU

↓

Results in your storage

No data leaves this boundary

Two editions, one platform

The same recognition engines, the same accuracy, the same output formats — scaled to your needs.

Workstation

Single server, one team

Everything runs on one machine with Docker Compose. Set up in an afternoon, run it from your desk. Ideal when one team needs local recognition capability without enterprise infrastructure.

Single NVIDIA GPU (RTX 3060 or better)
Web dashboard for upload, monitoring, and review
Docker Compose deployment — no Kubernetes needed
Custom model training on your own documents
All output formats: PageXML, ALTO, PDF, plain text

Best forResearch groups, small archives, individual departments

Enterprise

Kubernetes cluster, organisation-wide

Scales across multiple GPUs and worker nodes. Built for high-volume digitisation with REST API integration, job queuing, and monitoring. Deploys via Helm on Kubernetes or OpenShift.

Multiple GPUs with horizontal scaling and MIG support
REST API for programmatic submission and integration
Job queue with automatic distribution across workers
Prometheus monitoring and Grafana dashboards
Helm deployment with ArgoCD / GitOps support

Best forNational archives, large-scale digitisation programmes, institutions with IT infrastructure

Both editions include the full recognition engine catalog, model training, and all output formats.

Detailed system requirements

Platform

The Transkribus platform runs on your infrastructure as a set of containerised services. You get the full recognition stack — the same engines, models, and output formats used in the cloud version.

Text recognition (Standard HTR & Super Models)

Layout analysis & baseline detection

Table & field extraction

Named entity recognition

Custom model training

Web dashboard & REST API

Batch processing & job queue

All export formats (PageXML, ALTO, PDF, plain text)

Services — configured to your needs

Consulting & sizing

We review your documents, volumes, and infrastructure to recommend the right setup. Single workstation or Kubernetes cluster, one GPU or eight — depends on what you need.