Skip to content

German-OCR-3 v0.2 — Compact German Vision-OCR (Apr 2026)

Latest

Choose a tag to compare

@Keyvanhardani Keyvanhardani released this 23 Apr 13:12
· 1 commit to main since this release

🇩🇪📄 German-OCR-3 v0.2 ist da!

Kompakte, lokal lauffähige Vision-OCR für deutsche Geschäftsdokumente. Aus dem Bild kommt strikt validiertes JSON — ohne Cloud, ohne Vendor-Lock-in.

In einer Zeile installiert:

ollama pull Keyvan/german-ocr-3

Highlights auf 200+ echten anonymisierten DE-Rechnungen:

  • 100 % gültiges JSON
  • 95 % Sender korrekt erkannt
  • 0 % Halluzination
  • 2.7 GB — läuft auf 4–6 GB VRAM, Nano sogar CPU-only

Zwei Editionen, beide Apache 2.0:

  • Keyvan/german-ocr-nano (1.0 GB)
  • Keyvan/german-ocr-3 (2.7 GB, empfohlen)

Auf Hugging Face als GGUF — direkt nutzbar via Ollama, llama.cpp und vLLM. Komplett offen für die Community.

Wer lieber gemanagt mit größeren Modellen und EU-Datenhaltung will: german-ocr.de — hosted API, keine eigene Hardware.

👉 Hugging Face: https://huggingface.co/Keyven/german-ocr-3
👉 Ollama: https://ollama.com/Keyvan/german-ocr-3
👉 GitHub: https://github.com/Keyvanhardani/German-OCR
👉 Hosted API: https://german-ocr.de

Feedback & ⭐ willkommen!

#OCR #GermanAI #Python #VisionLanguageModel #OpenSource #GermanOCR #DocumentAI #LocalLLM #HuggingFace #Ollama #NLP #Apache2 #DSGVO #EUAIAct

Bild-Quelle: Gemini