Qwen 3.5, Gemma 4 battle for handwriting OCR crown

The tradeoff between accuracy and reliability is the new frontier for local multimodal models in industrial-scale OCR. Qwen 3.5 9B's "magic" quantization on Ollama provides near 100% accuracy but fails on 20% of pages due to flood-filling repetition loops. Gemma 4, launched April 3, 2026, offers 95% accuracy with a 30% speed increase and rock-solid reliability, making it the superior candidate for massive batch processing. The 50 million page scale represents a colossal compute challenge requiring a highly stable pipeline before full-scale deployment. The repetition loop issue in Qwen 3.5 appears to be a quantization artifact that repetition penalties fail to suppress in complex visual tasks, while local deployment remains the preferred path for sensitive document processing via Ollama GGUF models.