๐Ÿ‡ฉ๐Ÿ‡ช Rechly-DE-3B: 100% Local German Accounting & E-Rechnung AI

Ollama Model Hugging Face GitHub Evaluation Suite License Privacy by Design

The specialized 3-Billion parameter Small Language Model (SLM) for German Invoicing, the 2025/2026 E-Rechnung Mandate (EN 16931, XRechnung 3.0, ZUGFeRD 2.3), GoBD Pre-Audit, and DATEV SKR03/04 Bookkeeping.

Runs 100% on-device on Apple Silicon MacBooks, lightweight Linux VPS, or standard CPUs with zero cloud dependencies and < 2.5 GB RAM.


โšก 10-Second Quickstart

1. Run instantly with Ollama:

ollama run jonprompter/rechly-de-3b

2. Test an invoice in one line:

ollama run jonprompter/rechly-de-3b "Rechnung an Kanzlei Schmidt: 10 Std. IT-Forensik ร  140โ‚ฌ und 1x Anfahrt Pauschal 80โ‚ฌ, zahlbar in 14 Tagen."

3. Python 4-Liner:

import requests

resp = requests.post("http://localhost:11434/api/chat", json={
    "model": "jonprompter/rechly-de-3b",
    "messages": [{"role": "user", "content": "Kategorisiere fรผr SKR03 und SKR04: 'Geschรคftsessen im Restaurant 140โ‚ฌ mit Kunde'"}],
    "stream": False,
    "options": {"temperature": 0.0}
})
print(resp.json()["message"]["content"])
# Output: {"category": "Bewirtung", "skr03_account": "4650", "skr04_account": "6640", "deductible_percent": 70, "vat_rate": 19}

๐Ÿ›๏ธ The Rechly Model Family: 3 Tiers for Every Infrastructure

Model Tier Base Architecture Format / Quant Model Size RAM Footprint Inference Latency Target Deployment
Rechly-DE-3B (Flagship) Qwen2.5-3B-Instruct GGUF Q4_K_M 1.84 GB ~2.4 GB RAM ~1.5s / query Desktop, M1/M2/M3/M4 Mac, Production Server
Rechly-DE-1.5B (Slim) Qwen2.5-1.5B-Instruct GGUF Q4_K_M 940 MB ~1.0 GB RAM ~0.7s / query 1 GB VPS, Micro-Instances, Cloud Run, CI/CD
Rechly-DE-0.5B (Nano) Qwen2.5-0.5B-Instruct GGUF Q4_K_M 379 MB ~500 MB RAM ~0.3s / query Raspberry Pi 4/5, Edge Devices, Embedded IoT
# Pull & Run Flagship (3B)
ollama run jonprompter/rechly-de-3b

# Run Slim (1.5B) - 1 GB VPS
ollama create rechly-de-1.5b -f Modelfile.1.5b
ollama run rechly-de-1.5b

# Run Nano (0.5B) - 500 MB RAM / Raspberry Pi
ollama create rechly-de-0.5b -f Modelfile.0.5b
ollama run rechly-de-0.5b

๐Ÿšจ Why Rechly-DE-3B? The German E-Rechnung 2025/2026 Challenge

Under the Wachstumschancengesetz (ยง 14 UStG), traditional paper invoices and standard unstructured PDFs are legally obsolete for domestic B2B transactions. Over 3.5 million German companies are legally mandated to process structured E-Rechnungen (EN 16931, XRechnung 3.0, ZUGFeRD 2.3).

The Challenge with Cloud AI Services:

  • Privacy & Data Governance Concerns: Invoices, client names, hourly rates, and bank accounts contain sensitive personal and financial data. Processing them on third-party cloud servers requires Data Processing Agreements (DPA / AVV) and careful risk assessment under GDPR / DSGVO.
  • Token Costs: Processing verbose XML structures, long invoices, and receipt scans can lead to significant ongoing API operational costs.
  • Domain Accuracy on German Tax Law: General-purpose foundation models frequently misapply German statutory rules, such as confusing DATEV account numbers, misapplying the 70% entertainment expense deductibility cap (ยง 4 Abs. 5 Nr. 2 EStG), or omitting statutory tax-exemption notices (ยง 19 UStG, ยง 13b UStG).

The Rechly-DE-3B Architecture:

  • 100% Local & Offline (Zero Data Egress): Operates entirely on-device, preventing external data transmission and supporting the principle of Data Protection by Design (Art. 25 GDPR).
  • Specialized Accounting Adaptation: Fine-tuned on curated statutory German tax law and 30+ commonly used DATEV SKR03/SKR04 accounts.
  • Lightweight Footprint: 1.93 GB GGUF Q4_K_M running at 60โ€“80+ tokens/sec on Apple Silicon or standard CPU with < 2.5 GB RAM.
  • Formal Grammars: Ships with pre-configured GBNF grammars (grammars/*.gbnf) for mathematically constrained JSON schema adherence.

๐Ÿ“Š Internal Synthetic Evaluation Suite: Rechly-DE-3B vs. General LLMs

Rechly-DE-3B was evaluated against base foundation models using an internal synthetic test suite of 77 reproducible test vectors covering German tax provisions, DATEV account assignment, and formal JSON formatting. (Note: This is an internal project test suite and does not represent an external regulatory certification).

Evaluation Domain Qwen2.5-3B (Base) Llama 3 (8B) Mistral (7B) Rechly-DE-3B (Ours)
Model Size / RAM 3B (~2.5 GB) 8B (~6.0 GB) 7B (~5.5 GB) 3B (< 2.5 GB RAM)
Local Offline Inference โœ… Yes โœ… Yes โœ… Yes โœ… Yes (Zero Egress)
GoBD Tax Accuracy (ยง 19, ยง 13b, ยง 4, ยง 35a) 25.0% 58.0% 62.0% 100.0% (77/77 internal tests)
DATEV SKR03 & SKR04 Mapping 10.0% (Hallucinates) 45.0% 40.0% 100.0% (30+ Verified Accounts)
E-Rechnung EN 16931 / Leitweg-ID 15.0% (Invalid BT-10) 50.0% 48.0% 100.0% (XRechnung 3.0 & ZUGFeRD)
ZUGFeRD 2.3 PDF/A-3 Hybrid Output โŒ No โŒ No โŒ No โœ… Built-in Generator Script
Multimodal Receipt OCR Pipeline โŒ No โŒ No โŒ No โœ… Integrated Pipeline (Qwen2.5-VL)
DATEV EXTF Format 700 Batch Exporter โŒ No โŒ No โŒ No โœ… Conforming CSV & DUO ZIP Export
Internal Synthetic Test Suite Score 22.0% (17/77) 54.5% (42/77) 58.4% (45/77) 100.0% (77/77 Passed)

๐ŸŽฏ 10 Core Capabilities & Built-in Modules

1. Text-to-Invoice (Strict JSON Extraction)

Converts unstructured German voice notes, emails, or messages into structured, validated JSON:

  • Extracts client names, quantities, units, prices, due dates, and tax rates (19%, 7%, 0%).
  • Automatically detects tax exemptions: ยง 19 UStG (Kleinunternehmer), ยง 13b UStG (Reverse Charge), ยง 4 Nr. 1b (EU B2B), and ยง 4 Nr. 1a (Drittland Export).

2. ZUGFeRD 2.3 & Factur-X PDF/A-3 Hybrid Invoice Generator

Generates ready-to-send hybrid electronic invoices matching the 2025/2026 German mandate:

  • Combines a beautiful, human-readable visual PDF with an embedded, machine-readable factur-x.xml (EN 16931).
  • Associated file (/AF) registration and ISO 19005-3 compliant XMP metadata.
  • Generated via scripts/zugferd_generator.py or the /api/zugferd-generate endpoint.

3. Multimodal Receipt OCR to DATEV Bookings (Vision SLM)

Turns photos of crumpled receipts, restaurant slips (Bewirtungsbelege), fuel tickets (Tankbons), and scanned invoices directly into accounting entries:

  • Preprocesses and cleans receipt images with Pillow.
  • Extracts line items, merchant name, date, tax rate, and totals via local vision SLM (e.g., Qwen2.5-VL 3B).
  • Directly outputs ready-to-import DATEV EXTF CSV rows and GoBD pre-audit risk evaluations via scripts/receipt_vision_ocr.py.

4. DATEV SKR03 & SKR04 Expense Accounts (30+ Standard Accounts)

Maps natural language expense descriptions to standard German 4-digit chart accounts with statutory deductibility percentages:

  • Bewirtung (Restaurants): SKR03 4650, SKR04 6640 (70% abzugsfรคhig gem. ยง 4 Abs. 5 Nr. 2 EStG, 100% Vorsteuer)
  • SaaS & Cloud Software: SKR03 4964, SKR04 6837 (100% Betriebsausgabe)
  • GWG Geringwertige Wirtschaftsgรผter (< 800โ‚ฌ): SKR03 0480, SKR04 0670
  • Treibstoff & Kfz-Kosten: SKR03 4530, SKR04 6530
  • Telekommunikation & Internet: SKR03 4920, SKR04 6805
  • Reisekosten Fahrt (Bahn 7% MwSt): SKR03 4673, SKR04 6673
  • Reisekosten Hotel (รœbernachtung 7% MwSt): SKR03 4660, SKR04 6660
  • Fachliteratur (7% MwSt): SKR03 4940, SKR04 6815
  • Bankgebรผhren (0% steuerfrei): SKR03 4970 / 2100, SKR04 6855 / 7300

5. GoBD ยง 14 Abs. 4 UStG Pre-Audit Inspector

Pre-screens incoming supplier invoices for all 10 statutory mandatory fields to prevent tax auditors from disallowing input VAT deduction (Vorsteuerabzug):

  • Flags missing Steuernummer/USt-IdNr, missing consecutive invoice numbers, and missing delivery periods (Leistungszeitpunkt).
  • Outputs risk categories: KEIN_RISIKO, MITTEL, HOCH with legal remediation instructions.

6. DATEV Buchungssatz CSV Generator (EXTF)

Generates ready-to-import semicolon-separated DATEV EXTF ASCII lines for automated batch import into DATEV Rechnungswesen and Kanzlei-Rechnungswesen:

142,80;S;EUR;;;;4964;1200;9;1502;RE-2025-081;;;Google Workspace SaaS

7. B2B Line Item Rewriter

Transforms informal craftsman or developer shorthand into audit-proof, GoBD-compliant commercial invoice lines:

  • Input: "klo repariert und dichtung gewechselt"
  • Output: "Fachgerechte Sanitรคr-Instandsetzung: Austausch schadhafter Dichtungen, mechanische Rohrreinigung sowie abschlieรŸende Dichtheitsprรผfung."

8. Multi-Stage Mahnwesen (Dunning Letters)

Drafts legally disciplined payment reminders across German commercial stages:

  • Stufe 1: Freundliche Zahlungserinnerung (diplomatisch, ohne Mahngebรผhren)
  • Stufe 2: 1. Mahnung (formell mit 7-Tage-Nachfrist)
  • Stufe 3: Letzte Mahnung (bestimmt, Ankรผndigung gerichtlicher Schritte und Inkassokosten)

9. DATEV EXTF 700 Batch & DUO ZIP Export Utility

Exports batches of booking entries according to the DATEV Format 700 specification (independent community implementation):

  • 27-to-31-column semicolon format with standard BU-Schlรผssel (9, 8, 3, 2, 94).
  • DATEV Unternehmen online (DUO) compatible ZIP structure containing EXTF_Buchungsstapel.csv and XML manifest (document.xml).
  • Dual-encoding support (CP1252 and UTF-8-SIG) for compatibility with DATEV Kanzlei-Rechnungswesen import tools.

10. Mathematical Reconciler & ยง 35a Handwerkerleistungen Splitter

Applies commercial rounding (ยง 35 UStDV) and statutory tax reductions:

  • Reconciles mixed-rate tax groupings (19%, 7%, 0%) with cent-exact arithmetic.
  • Automatically isolates craftsman labor and travel costs from materials under ยง 35a Abs. 3 EStG, calculating the eligible 20% direct tax reduction (capped at 1.200 โ‚ฌ p.a.).

๐Ÿ”ฌ DPO / ORPO Preference Alignment Dataset

To reduce legal hallucinations on German accounting rules, Rechly-DE-3B includes a curated preference alignment dataset (data/rechly_preference_pairs.jsonl) with 101 (prompt, chosen, rejected) pairs.

  • Negative samples (rejected) capture critical real-world failure modes:
    • Charging 19% VAT on Kleinunternehmer invoices (violating ยง 19 UStG and triggering ยง 14c tax liabilities).
    • Claiming 100% tax deduction on client entertainment (violating ยง 4 Abs. 5 Nr. 2 EStG).
    • Approving vendor invoices without a valid Steuernummer.
    • Inventing non-existent 4-digit DATEV codes.
  • Training Harness (scripts/train_orpo.py): Single-stage ORPO alignment via Hugging Face TRL and PEFT/QLoRA.

๐Ÿ”’ GBNF Grammars: Formal Schema Constraints

In financial production systems, JSON schema validation cannot rely on unconstrained generation. Rechly-DE-3B includes formal GBNF (GGML BNF) grammar files:


๐Ÿ’ป Integration Examples

Python (OpenAI-compatible)

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:11434/v1",
    api_key="ollama"  # not needed locally
)

response = client.chat.completions.create(
    model="jonprompter/rechly-de-3b",
    messages=[
        {"role": "user", "content": "Optimiere diese Tรคtigkeitsbeschreibung fรผr die Rechnungsstellung: 'homepage schneller gemacht und alte bilder gelรถscht'"}
    ],
    temperature=0.0
)
print(response.choices[0].message.content)

cURL

curl -X POST http://localhost:11434/api/chat -d '{
  "model": "jonprompter/rechly-de-3b",
  "messages": [{"role": "user", "content": "Erstelle eine Mahnung Stufe 1 an Kanzlei Schmidt รผber 1.200โ‚ฌ seit 10 Tagen fรคllig."}],
  "stream": false,
  "options": {"temperature": 0.0}
}'

๐Ÿ‘ฅ Who Is This For?

  1. Developers & SaaS Founders: Build automated German invoicing and E-Rechnung features into your CRM, ERP, or SaaS without paying OpenAI API fees.
  2. Freelancers & Small Businesses (Selbststรคndige / KMU): Automate invoice drafting and expense categorization on your own local device without third-party cloud data exposure.
  3. Tax Advisors & Bookkeepers (Steuerberater / Kanzleien): Speed up receipt pre-sorting and generate DATEV EXTF-compatible semicolon CSV files ready for import review.
  4. Enterprises & Privacy-Conscious Organizations: Run specialized accounting assistants on air-gapped on-premise infrastructure with zero risk of external cloud data egress.

โš–๏ธ Legal, GDPR & Benchmark Disclaimers

1. No Tax Advice (Keine Steuerberatung)

Rechly-DE-3B provides automated text extraction and schema-constrained formatting based on public German statutory standards (UStG, GoBD, EStG). It is an open-source productivity tool and does not replace certified tax advice by an authorized German Steuerberater or certified tax professional (gem. StBerG / RDG).

2. GDPR / DSGVO & Data Protection by Design

*The model runs 100% locally on-device with zero external network egress to third-party AI APIs, supporting technical data confidentiality and the principle of Data Protection by Design & Default (Art. 25 GDPR / DSGVO). However, legal GDPR compliance is an organizational responsibility of the entity acting as the Data Controller (Art. 4(7) GDPR). Using local software does not automatically grant a certification of full organizational compliance (e.g. data retention policies, access permissions, and lawful basis under Art. 6 GDPR).*

3. Internal Synthetic Benchmark Notice

All benchmark scores (including the 77-case evaluation suite) represent internal, self-reported synthetic test suites developed within this repository to verify statutory edge cases and schema validity. They do not constitute an independent audit, official government benchmark, or certification by any statutory authority, standard-setting body, or accredited testing laboratory.

4. Trademarks & Third-Party Notice

DATEVยฎ is a registered trademark of DATEV eG. ZUGFeRD is a registered specification of the Forum elektronische Rechnung Deutschland (FeRD / AWV e.V.). Factur-X is a registered trademark of FNFE-MPE. This project is an independent open-source community toolkit and is not affiliated with, endorsed by, or certified by DATEV eG, FeRD, or FNFE-MPE.


๐Ÿ“œ Citation & Open Source

@misc{rechly_de_3b_2026,
  author = {Rechly Team and Contributors},
  title = {Rechly-DE-3B: A Specialized Open Small Language Model for German Invoicing, GoBD Compliance, E-Rechnung, and DATEV Bookkeeping},
  year = {2026},
  publisher = {Hugging Face and GitHub},
  howpublished = {\url{https://huggingface.co/jonprompter/rechly-de-3b}}
}

โญ If you find Rechly-DE-3B useful, please star the model on Hugging Face and GitHub to support independent German open-source AI!

Downloads last month
1,072
Safetensors
Model size
3B params
Tensor type
U32
ยท
F16
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ 1 Ask for provider support

Model tree for jonprompter/rechly-de-3b

Base model

Qwen/Qwen2.5-3B
Quantized
(340)
this model