Instructions to use jonprompter/rechly-de-3b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use jonprompter/rechly-de-3b with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("jonprompter/rechly-de-3b") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use jonprompter/rechly-de-3b with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf jonprompter/rechly-de-3b:Q4_K_M # Run inference directly in the terminal: llama cli -hf jonprompter/rechly-de-3b:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf jonprompter/rechly-de-3b:Q4_K_M # Run inference directly in the terminal: llama cli -hf jonprompter/rechly-de-3b:Q4_K_M
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf jonprompter/rechly-de-3b:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf jonprompter/rechly-de-3b:Q4_K_M
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf jonprompter/rechly-de-3b:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf jonprompter/rechly-de-3b:Q4_K_M
Use Docker
docker model run hf.co/jonprompter/rechly-de-3b:Q4_K_M
- Llama new
- LM Studio
- Jan
- vLLM
How to use jonprompter/rechly-de-3b with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "jonprompter/rechly-de-3b" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "jonprompter/rechly-de-3b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/jonprompter/rechly-de-3b:Q4_K_M
- Ollama
How to use jonprompter/rechly-de-3b with Ollama:
ollama run hf.co/jonprompter/rechly-de-3b:Q4_K_M
- Unsloth Desktop
- Pi
How to use jonprompter/rechly-de-3b with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "jonprompter/rechly-de-3b"
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "jonprompter/rechly-de-3b" } ] } } }Run Pi
# Start Pi in your project directory: pi
- MLX LM
How to use jonprompter/rechly-de-3b with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "jonprompter/rechly-de-3b"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "jonprompter/rechly-de-3b" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "jonprompter/rechly-de-3b", "messages": [ {"role": "user", "content": "Hello"} ] }' - Docker Model Runner
How to use jonprompter/rechly-de-3b with Docker Model Runner:
docker model run hf.co/jonprompter/rechly-de-3b:Q4_K_M
- Lemonade
How to use jonprompter/rechly-de-3b with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull jonprompter/rechly-de-3b:Q4_K_M
Run and chat with the model
lemonade run user.rechly-de-3b-Q4_K_M
List all available models
lemonade list
- Hermes Agent
How to use jonprompter/rechly-de-3b with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "jonprompter/rechly-de-3b"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default jonprompter/rechly-de-3b
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use jonprompter/rechly-de-3b with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "jonprompter/rechly-de-3b"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "jonprompter/rechly-de-3b" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
Rechly-DE-3B: 100% Local German Accounting AI Model 🇩🇪
Rechly-DE-3B is a specialized, open-weights 3-Billion parameter language model engineered specifically for German invoicing, GoBD tax compliance, and DATEV bookkeeping for Rechly and the German open-source AI ecosystem.
Trained on curated German statutory and bookkeeping datasets, Rechly-DE-3B runs 100% locally on standard consumer hardware (Apple Silicon M1–M4, CPUs, and lightweight Linux VPS), ensuring absolute DSGVO / GDPR compliance with zero data egress.
⚡ Key Highlights & Specifications
| Feature | Specification |
|---|---|
| Base Architecture | Qwen/Qwen2.5-3B-Instruct |
| Quantization | GGUF Q4_K_M (~1.84 GB file size) |
| RAM Footprint | < 2.5 GB RAM (Runs comfortably on MacBook Air or 4GB RAM VPS) |
| Inference Speed | 60–80+ tokens/sec via Ollama on Apple Silicon Metal |
| Fine-Tuning Method | QLoRA (MLX on Apple Silicon Metal, PyTorch PEFT, Unsloth on Colab T4) |
| Evaluation Score | 95.6% (43/45 Passed, Grade A+) on Comprehensive Audit Benchmark |
| Gateway Guarantee | 100.0% Pass Rate via FastAPI deterministic schema validator |
| Compliance | German GoBD, UStG (§ 19, § 13b, § 4 Nr. 1b, § 25a), DATEV SKR03 / SKR04 |
| License | Apache 2.0 |
🎯 5 Core Accounting Capabilities
Rechly-DE-3B is specifically fine-tuned for five distinct enterprise bookkeeping tasks:
1. Text-to-Invoice (Strict JSON Extraction)
Extracts client name, due days, payment terms, and line items (with quantities, units, prices, and tax rates 19%, 7%, 0%) from unstructured German text into a strict, validated JSON schema.
Input:
Rechnung an Kanzlei Schmidt: 12 Stunden IT-Beratung à 120€ und 1x Server-Setup Pauschal 250€, zahlbar in 14 Tagen.
Output:
{
"client_name": "Kanzlei Schmidt",
"due_days": 14,
"payment_terms": "14 Tage",
"notes": "",
"items": [
{
"description": "IT-Beratung",
"quantity": 12,
"unit_of_measure": "Stunde",
"price": 120.0,
"tax_rate_percent": 19
},
{
"description": "Server-Setup",
"quantity": 1,
"unit_of_measure": "Pauschal",
"price": 250.0,
"tax_rate_percent": 19
}
]
}
2. B2B Line Item Rewriter
Converts raw, informal trade and service notes into audit-proof, GoBD-compliant descriptions suitable for professional B2B invoices.
- Input:
"website schneller gemacht und datenbank geputzt" - Output:
"Performance-Optimierung der Webpräsenz: Bereinigung und Indexierung relationaler Datenbanktabellen sowie messbare Ladezeitreduktion"
3. German Tax & GoBD Statutory Clauses
Generates exact, legally mandated statutory clauses under German tax law:
- § 19 UStG (Kleinunternehmer):
"Gemäß § 19 UStG wird keine Umsatzsteuer berechnet (Kleinunternehmerregelung)." - § 13b UStG (Reverse Charge / Bauleistungen):
"Steuerschuldnerschaft des Leistungsempfängers (Reverse Charge gem. § 13b Abs. 2 Nr. 4 i.V.m. Abs. 5 UStG)." - Innergemeinschaftliche Lieferung EU:
"Steuerfreie innergemeinschaftliche Lieferung gem. § 4 Nr. 1b i.V.m. § 6a UStG." - Differenzbesteuerung:
"Sonderregelung für Gebrauchtgegenstände gem. § 25a UStG (Differenzbesteuerung)." - Aufbewahrungsfristen (§ 147 AO):
"Gemäß § 147 AO und den GoBD müssen Buchungsbelege und Rechnungen 10 Jahre aufbewahrt werden; Handels- und Geschäftsbriefe 6 Jahre."
4. Mahnwesen (Multi-Tiered Payment Reminders)
Produces tone-appropriate German dunning letters across 3 standard stages:
- Stufe 1: Freundliche Zahlungserinnerung (sachlich, partnerschaftlich, keine Mahngebühren)
- Stufe 2: 1. Mahnung (formell mit 7-Tage-Nachfrist)
- Stufe 3: Letzte Mahnung (bestimmt, Ankündigung gerichtlicher Schritte / Inkasso und Mahnkosten)
5. DATEV SKR03 / SKR04 Expense Categorization
Maps receipt text to standard German standard chart of accounts (SKR03 and SKR04), specifies the Vorsteuer rate (19% / 7% / 0%), and indicates corporate tax deductibility percentages (e.g. 70% Bewirtung nach § 4 Abs. 5 EStG, 100% SaaS/Software).
📊 Comprehensive 45-Point Audit Benchmark
Evaluated against the rigorous 45-point German accounting audit battery (scripts/benchmark_comprehensive.py):
| Domain | Base Qwen 2.5 3B | Rechly-DE-3B (Raw LLM) | With Rechly API Gateway |
|---|---|---|---|
| Domain 1: Text-to-Invoice (Strict JSON) | 20.0% (schema drift, syntax errors) | 80.0% (8/10) | 100.0% (validated) |
| Domain 2: B2B Line Item Rewriter | 25.0% (generic English/colloquial) | 100.0% (8/8) | 100.0% |
| Domain 3: Tax & GoBD Clauses | 40.0% (missing UStG citations) | 100.0% (10/10) | 100.0% |
| Domain 4: Mahnwesen Dunning Letters | 28.5% (hallucinated JSON / informal) | 100.0% (7/7) | 100.0% |
| Domain 5: DATEV SKR03/04 Accounts | 10.0% (invalid codes / wrong tax) | 100.0% (10/10) | 100.0% |
| Overall Scorecard | 25.0% (Grade F) | 95.6% (43/45, Grade A+) | 100.0% (Grade A+) |
- Average Inference Latency: 2.52 seconds per query on Apple Silicon M-series (Metal GPU)
- Memory Footprint: 2.02 GB VRAM / RAM (Q4_K_M GGUF)
🚀 Quickstart & Usage
1. Ollama (Recommended)
# Register locally
ollama create rechly-de-3b -f Modelfile
# Run interactive query
ollama run rechly-de-3b "Rechnung an WebTech GmbH: 10 Stunden Programmierung à 95€"
2. Python Client SDK
from rechly_client import RechlyClient
client = RechlyClient()
# 1. Text-to-Invoice
invoice = client.text_to_invoice("Rechnung an Kanzlei Schmidt: 5 Std. Beratung à 120€")
# 2. DATEV Expense Categorization
expense = client.categorize_expense("Geschäftsessen mit Kunde im Restaurant 140€")
print(expense["skr03_account"]) # 4650
print(expense["deductible_percent"]) # 70
3. Apple Silicon Native (MLX)
from mlx_lm import load, generate
model, tokenizer = load("rechly-de-3b-mlx")
prompt = "<|im_start|>user\nKategorisiere für SKR03: 'Google Workspace Cloud 70€'<|im_end|>\n<|im_start|>assistant\n"
response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(response)
⚖️ Legal & Tax Compliance Disclaimer
Rechly-DE-3B provides automated assistance based on German tax standards (GoBD, UStG, DATEV). It is intended as an assistive productivity tool for invoicing and bookkeeping and does not replace certified tax advice by an authorized Steuerberater.
📜 Citation
@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, and DATEV Bookkeeping},
year = {2026},
publisher = {GitHub and Hugging Face},
howpublished = {\url{https://github.com/rechly/rechly_llm}}
}
- Downloads last month
- 213