odooclaw-vision

odooclaw-vision

Vision model for OdooClaw — on-premise document extraction (invoices, delivery notes) with no cloud dependencies.

Base model: zai-org/GLM-OCR (0.9B params, MIT license) — the best quality-per-parameter OCR of 2026, exceptional at tables and structured documents. GGUF conversion by ggml-org.

Part of the OdooClaw collection. MLX version for Apple Silicon: odooclaw-vision-mlx.

Evaluation

Files

Usage with llama.cpp

llama-server \
  -m odooclaw-vision-Q5_K_M.gguf \
  --mmproj mmproj-odooclaw-vision-Q8_0.gguf \
  --host 0.0.0.0 --port 8093 \
  -c 8192 --parallel 1 --temp 0.0 \
  --alias odooclaw-vision

OpenAI-compatible endpoint:

curl http://localhost:8093/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "odooclaw-vision",
    "messages": [{"role": "user", "content": [
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
      {"type": "text", "text": "Extract all text from this invoice document, preserving table structure."}
    ]}],
    "temperature": 0,
    "max_tokens": 2048
  }'

OdooClaw pipeline architecture

Invoice PDF → odooclaw-vision (image → structured text)
            → odooclaw-light (text → JSON: partner, vat, ref, date, total, lines)
            → business rules (validation: reverse charge, currency, sanity checks)
            → account_dynamic_rules (Odoo rules: account, analytics, taxes)
            → vendor bill in Odoo

Performance

• CPU (N100, 4 cores): ~10-40s per invoice (Q5_K_M)
• Strix Halo (Vulkan): 0.2-0.5s per invoice
• Apple Silicon (MLX): 0.21s load + generate (see odooclaw-vision-mlx)
• Quality: totals, partners, dates and line items extracted correctly from real production invoices (SIEPER, utilities, telecom)

License

MIT. Commercial use allowed.