Dokumentli® – The AI model that beats LLMs at documents.
Dokumentli® reads, classifies, and structures business documents more precisely and faster than general-purpose frontier models – and can run fully air-gapped on a single consumer GPU with 24 GB VRAM.
Runs on 1 GPU with 24 GB. Beats seven models at documents.
Built for European enterprises and regulated industries with high, recurring document volumes – from financial services and insurance to construction, trade, and logistics.
Avg. accuracy ([email protected]), highest score among 7 models tested
faster on average than the most accurate of the six reference models
Deployment options, from on-prem to API
of GPU memory are enough for full air-gapped operation on a single consumer graphics card
Air-gapped capable
Fully on-prem operation on a single consumer GPU with 24 GB VRAM – no data center, no cloud connection required.
What Is a Small Special-Purpose Vision Language Model?
A Vision Language Model (VLM) processes image and text together. It reads a document page directly as an image – including layout, tables, stamps, and handwriting – and answers a text-based query about it. No separate OCR pre-processing step is required.
Image + text in one model
Layout, tables, and text content are interpreted together, without a separate OCR pipeline.
Specialized for documents
Trained on real business documents instead of general web text.
Compact
Fewer parameters and lower compute per document – runs entirely on a single consumer GPU with 24 GB VRAM.
is enough for full air-gapped operation – no data center, no cloud connection needed.
of VRAM on a single consumer graphics card are enough to run the entire model.
Highest Accuracy at the Highest Speed
In an independent test across 5 benchmark datasets comprising a total of 3,368 real business documents, Dokumentli® was tested against six current reference models: Claude Opus 4.8, GPT-5.6 Sol, Gemini 3.6 Flash, Gemini 3.7 Flash, Qwen3.6-Plus, and the open model Gemma 4 26B-A4B-IT. The quality metric is [email protected] (Average Normalized Levenshtein Similarity with a threshold of 0.8), a common variant of the ANLS industry standard for evaluating document extraction.
5 / 5
Benchmarks in which Dokumentli® outperforms Gemini 3.6 Flash, the strongest of the six reference models.
7
models tested in total: Dokumentli® plus six general-purpose frontier models as reference.
For details, see the Benchmark Report
Results by industry and model, confidence intervals, positioning matrices, and the full methodology are documented in the separate Parashift Dokumentli® Benchmark Report .
Why a Specialized Model Beats General-Purpose LLMs
General-purpose frontier models are trained to cover a very broad range of tasks. For the narrowly defined task of "read this business document and extract these fields," that training objective misses the actual task. Four factors explain why a smaller, specialized model comes out ahead in these benchmarks:
Millions of real business documents instead of general web text
Including the quirks of invoices, freight documents, and contracts from regulated industries.
Image and text are processed in a single step
No lossy OCR pre-processing step – relevant for tables, multi-column layouts, stamps, and handwritten notes.
A smaller context budget per request
This noticeably reduces latency and compute per document – and is what makes operation on a single consumer GPU with 24 GB VRAM possible in the first place.
Continuous fine-tuning instead of just prompt adjustments
Fine-tuning on company-specific templates and edge cases, instead of just generic prompt adjustments.
Generalist (general-purpose LLM)
One model, many possible tasks: conversation, code, world knowledge, documents. Document understanding is one capability among many, not the training focus.
Specialist (Dokumentli®)
One model, one task: reading, classifying, and structuring business documents. Training data, architecture, and context budget are fully aligned to it.
What Dokumentli® Can't Do
The specialization that makes Dokumentli® strong at document intelligence tasks is also a deliberate boundary: Dokumentli® is not a general-purpose assistant. Outside document processing, a general-purpose model — or better yet, a model specialized for the specific automation task — is the right choice.
| Task | Dokumentli® | General-purpose frontier model |
|---|---|---|
| Classify Documents & Extract Fields | ✓ Core task, trained for it | ✕ Possiblewith lower accuracy and speed, higher costs, and greater security/compliance overhead |
| Free-Form Reasoning & Open Conversation | ✕ Not intended | ✓ Core task |
| Programming / Code Generation | ✕ Not intended | ✓ Core task |
| General knowledge questions outside document content | ✕ Not intended | ✓ Core task |
| Free-Form Creative Writing | ✕ Not intended | ✓ Core task |
| Multi-step agentic tasks unrelated to documents | ✕ Not intended | ✓ Possible |
Use Dokumentli® When…
- high volumes of business documents need to be classified or have fields extracted from them
- data sovereignty, air-gapped operation, or predictable costs matter
- structured, schema-compliant output is needed instead of free text
Use a General-Purpose LLM When…
- the task involves open-ended reasoning, conversation, or code
- general world knowledge outside the document content is required
- it involves free-form creative writing or agentic tasks unrelated to documents
Combinable, Not Exclusive
Companies use Dokumentli® specifically for inbound document processing and combine it, where needed, with general-purpose models for downstream steps unrelated to documents. Both can be orchestrated through your own workflow solution.
Four Ways to Run Dokumentli®
Depending on your data sovereignty, compliance, and integration requirements, Dokumentli® can be run in four ways — from fully isolated operation in your own data center to direct API integration.
The Model and the Node Component
Dokumentli® is the AI model itself. For easy deployment on your own hardware, Parashift additionally provides the Dokumentli® Node, a lightweight component that packages the model and enables on-prem or air-gapped setup in a short time, without having to build your own deployment setup.
On-Prem / Own Hardware
Maximum Data Sovereignty- Location: your own data center, via Dokumentli® Node
- Air-gapped capable: no data leaves your own network
- Hardware: one consumer-grade GPU with 24 GB VRAM is enough
- Typical for: the highest compliance requirements, no internet connection needed
Private Cloud Instance
Dedicated Tenant- Location: isolated cloud instance in your region/zone of choice
- Data sovereignty: very high, dedicated tenant
- Hardware: provided by the cloud provider, not by Parashift
- Typical for: regulated enterprises with their own cloud strategy
Parashift Plattform
Coming soon- Document Intelligence AI Guardrail Platform
- Location: Parashift's European sovereign cloud platform
- Data sovereignty: high, compliance zones EU, Switzerland, Germany
- Additional features: Ingestion, separation, validation, automation, 30+ connectors
- Typical for: teams who want to use the entire document pipeline
Dokumentli® API
Coming soon- Direct model cloud API access
- Location: programmatic endpoint, no platform UI
- Hardware: none, fully managed
- Flexible: against your own Dokumentli Node or a Parashift endpoint
- Typical for: teams embedding Dokumentli® into their own application
Open to any agentic automation and workflow solution
On-prem, private cloud instance, and the Dokumentli® API can be connected to practically any document or workflow automation solution on the market — whether RPA, iPaaS, ERP, DMS, or a custom-built application. The only requirement is that the solution can make an API call; no binding to a specific platform is required.
Fits into Your Existing Stack or the Parashift Platform
Dokumentli® is deliberately built to fit into virtually any existing document or workflow automation landscape, regardless of the deployment option chosen. In addition, Dokumentli® can be integrated in two fundamental ways.
Programmatic, into any existing solution
- Full control: Dokumentli® as a building block in your ERP, DMS, RPA/iPaaS solution, or your own application.
- Your own schema: the request defines the desired data format.
- Vendor-agnostic: no binding to a specific platform, just one API call needed.
Additionally orchestrated via the Parashift Platform
- For anyone who doesn't want to run their own orchestration: a built-in validation interface, human-in-the-loop without building your own frontend.
- 30+ connectors: ERP, DMS, CRM systems, and email, without your own middleware.
- Governance included: Confidence scores, routing thresholds, audit trail, 2-/3-way match.
In practice, this isn't an either-or decision: existing document and workflow automation solutions integrate Dokumentli® directly via the API, while teams without their own orchestration can additionally use the ready-made Parashift Platform. Both patterns can be combined on the same Dokumentli® instance.
RPA · iPaaS · ERP · DMS · CRM · Email inboxes · Custom applications
Why Dokumentli® Is More Cost-Effective Than General-Purpose LLM APIs
General-purpose frontier models at cloud providers are typically billed per API call and per token processed. For companies with high and fluctuating document volumes, this creates two structural disadvantages: costs scale directly with volume, with no predictable ceiling, and additionally with the context budget a general-purpose model requires per request.
Cost trend over time, schematic illustration
Illustrative, without specific amounts: pay-per-call costs fluctuate with volume and demand spikes; a page-quota subscription stays constant and predictable.
Pay-per-call
General-purpose LLM APIs: costs rise directly and without a ceiling as volume and token usage increase.
Capacity-Based
Dokumentli®: subscription with a page quota, predictable in advance, independent of demand spikes.
This can be objectively measured through the lower compute per document: Dokumentli® processes a document with noticeably less compute power than general-purpose reference models, while achieving higher accuracy – and can run fully on a single consumer GPU with 24 GB VRAM instead of cloud data centers. That makes costs fundamentally predictable, independent of any individual cloud provider's pricing model.
Subscriptions
Dokumentli® is offered in several subscription tiers, scaled by monthly document volume, from pilot projects to high-volume enterprise deployment. Each tier includes 4 base-model updates per year, so accuracy and feature improvements flow in automatically, with no new contract.
Trust That Goes Beyond Compliance
Data Sovereignty
Operation on-prem, air-gapped, or in dedicated compliance zones for the EU, Switzerland, and Germany – your document data never leaves the chosen zone.
No Shadow AI
Dedicated tenants instead of shared infrastructure: your confidential documents never flow into publicly accessible model training.
Governance & Audit Trail
Confidence scores, routing thresholds, and complete audit trails on the Parashift Platform, for traceable decisions at every process step.
A Specialized Model for a Clearly Defined Task
Dokumentli® achieves the highest average extraction accuracy among seven models tested, and is also the fastest model in the field on average – with significantly lower compute and the option to run fully air-gapped on a single consumer GPU with 24 GB VRAM.
More accurate
than general-purpose LLMs
Faster
than general-purpose LLMs
More cost-effective
to operate
Air-gapped
on 1 GPU (24 GB)
What Automation Leads Ask Us First
Yes. In the on-prem / own hardware deployment option, Dokumentli® runs in your own data center via the Dokumentli® Node — no data leaves your own network, and a single consumer-grade GPU with 24 GB VRAM is enough. Neither an additional data center nor an internet connection is required.
In an independent test across 5 benchmark datasets comprising a total of 3,368 real business documents, against seven models including Dokumentli®, using [email protected] (Average Normalized Levenshtein Similarity with a threshold of 0.8) as the quality metric. The full methodology and detailed results are documented in the separate Benchmark Report.
Yes — combinable, not exclusive. Many companies use Dokumentli® specifically for inbound document processing and combine it, where needed, with general-purpose models for downstream steps unrelated to documents, such as open-ended reasoning or code.
Via a subscription with a page quota instead of pay-per-call, scaled by monthly document volume from pilot projects to high-volume deployment. Each tier includes 4 base-model updates per year, with no new contract.
Four: on-prem / own hardware and private cloud instance are available today; the Parashift Platform and the Dokumentli® API are coming soon. All four can be connected to practically any document or workflow automation solution.
Dokumentli® is not a general-purpose assistant: for open-ended reasoning, code generation, general knowledge questions, free-form creative writing, or multi-step agentic tasks unrelated to documents, a general-purpose model is the right choice.
Test Dokumentli® on your own documents.
30 seconds in the playground show more than any slide: upload one of your own documents and see how Dokumentli® reads, classifies, and structures it.
