Qwen3.5-4B nails local handwriting OCR

A 4B local multimodal model turning messy handwriting into clean structured text on consumer hardware is the real story here, even if this is still a field report rather than a rigorous benchmark.

• –The model did more than character recognition: it preserved sections, bullet hierarchy, labels, and arrows from a handwritten knowledge map
• –The reported setup was practical for hobbyists and builders: llama.cpp, an Unsloth Q4_K_XL GGUF, and roughly 46 tokens/sec on a 3070 laptop GPU
• –This is exactly the kind of multimodal capability that matters for document pipelines, personal knowledge capture, and OCR-plus-understanding workflows
• –The caveat is obvious: one Reddit sample is not a benchmark suite, so treat this as a strong signal of utility, not proof that Qwen3.5-4B now leads OCR overall