Qwen3.5 27B tops OCR redaction tests

Qwen3.5 27B is good enough to be useful, but not good enough to be trusted blindly. That’s the real story here: local redaction has moved from demo territory into a productive human-in-the-loop workflow.

• –Handwritten OCR is the clearest win, with better word capture and bounding boxes than smaller Qwen variants.
• –Custom-entity redaction is a strong fit because it leans on semantic span finding more than perfect page layout.
• –Dense pages still trigger skipped lines, so omissions remain the main failure mode.
• –Face redaction is still fragile because detection succeeds more often than full coverage.
• –The best workflow is hybrid: deterministic extraction first, PaddleOCR for easy text, Qwen3.5 27B for the hard leftovers.
• –At 4-bit quantization, the model finally lands in the consumer-GPU sweet spot that makes local redaction practical.