🇩🇪 Rechly-DE-1.5B (Slim): 100% Local German Accounting AI for 1 GB VPS
The ultra-lightweight 1.5-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 inside ~1.0 GB RAM on budget 1 GB VPS micro-instances, Cloud Run, or low-cost edge containers with sub-second latency.
🏛️ The Rechly Model Family
- ⚡ Rechly-DE-3B (Flagship) — 3B params · 1.84 GB · Desktop & Production Server
- 🪶 Rechly-DE-1.5B (Slim) — 1.5B params · 940 MB · 1 GB VPS & Cloud Run
- 🔬 Rechly-DE-0.5B (Nano) — 0.5B params · 379 MB · Raspberry Pi & 512MB RAM
⚡ 10-Second Quickstart
1. Run with Ollama:
# Clone and build local Ollama model
huggingface-cli download jonprompter/rechly-de-1.5b rechly-de-1.5b-q4_k_m.gguf Modelfile --local-dir .
ollama create rechly-de-1.5b -f Modelfile
ollama run rechly-de-1.5b
2. Test an invoice in one line:
ollama run rechly-de-1.5b "Rechnung an Kanzlei Schmidt: 10 Std. IT-Forensik à 140€ und 1x Anfahrt Pauschal 80€, zahlbar in 14 Tagen."
3. Python 4-Liner:
import requests
resp = requests.post("http://localhost:11434/api/chat", json={
"model": "rechly-de-1.5b",
"messages": [{"role": "user", "content": "Kategorisiere für SKR03 und SKR04: 'Geschäftsessen im Restaurant 140€ mit Kunde'"}],
"stream": False,
"options": {"temperature": 0.0}
})
print(resp.json()["message"]["content"])
⚡ Specifications
| Parameter | Specification |
|---|---|
| Base Architecture | Qwen/Qwen2.5-1.5B-Instruct (Apache 2.0 open-weights) |
| Quantization | GGUF Q4_K_M (~940 MB file size) |
| RAM Footprint | ~1.0 GB RAM (Runs comfortably on 1 GB VPS, Cloud Run, Micro Instances) |
| Inference Speed | 120+ tokens/sec on Apple Silicon Metal / 0.7s per query |
| Internal Test Suite | 100.0% Pass (77/77 internal tests) on synthetic E-Rechnung XRechnung 3.0 & DATEV SKR edge cases |
| Zero-Footprint Utilities | Hybrid ZUGFeRD 2.3 PDF/A-3 generator, DATEV EXTF 700 batch utility, and reconciler run deterministically in Python with 0 MB LLM memory overhead |
🎯 Supported Accounting Domains
- Text-to-Invoice (Strict JSON Extraction): Automatic detection of Kleinunternehmer (§ 19 UStG), Reverse Charge (§ 13b UStG, § 3a UStG), and Export (§ 4 Nr. 1a UStG).
- E-Rechnung EN 16931 / XRechnung 3.0: Validated BT-10 Leitweg-ID extraction and UBL 2.1 XML serialization.
- DATEV SKR03 / SKR04 Mapping: 30+ chart accounts with statutory 70% Bewirtung limits.
- GoBD § 14 Abs. 4 Pre-Audit: Pre-checks incoming invoices to assist in identifying missing mandatory items.
- Multi-Stage Mahnwesen: Dunning letters for Stufe 1, 2, and 3.
- B2B Line Item Rewriter: Formalizes informal notes into professional position texts.
⚖️ Legal, GDPR & Benchmark Disclaimers
1. No Tax Advice (Keine Steuerberatung)
Rechly-DE-1.5B provides automated text extraction and schema formatting based on public statutory standards. It is an open-source productivity tool and does not replace certified tax advice by an authorized Steuerberater (gem. StBerG / RDG).
2. GDPR / DSGVO Notice
*The model runs locally on-device with zero network egress, supporting Data Protection by Design & Default (Art. 25 GDPR). However, legal GDPR compliance remains the responsibility of the entity acting as the Data Controller (Art. 4(7) GDPR).*
3. Internal Synthetic Benchmark Notice
Reported benchmark figures represent internal, synthetic test suites developed within this project. They do not constitute an official government audit or certified third-party endorsement.
4. Trademarks & Third-Party Notice
DATEV® is a registered trademark of DATEV eG. ZUGFeRD is a registered specification of FeRD / AWV e.V. This project is an independent community project and is not affiliated with or endorsed by DATEV eG or FeRD.
🔒 License & Open Source
Licensed under the Apache 2.0 License. Built by the Rechly team for privacy-first, on-premise German accounting.