Qwen 3 VL — vision specialist

Most open vision-language models disappoint on real documents. Qwen 3 VL is the exception — dense charts, handwriting, multilingual signage, long PDFs. It does not match GPT-5 on every eval but it wins on price-per-page by a wide margin.

• Mixed-language OCR (Chinese + English on one page)
• Invoice and receipt parsing for finance ops
• Diagrams with annotations
• Handwritten notes

A doc-AI pipeline on Qwen 3 VL

1. 1PDF splitter produces page images at 300 DPI
2. 2Qwen 3 VL emits structured JSON per page
3. 3A downstream LLM validates and merges
4. 4Low-confidence pages route to human review

resp = Generation.call(
    model="qwen-vl-max",
    messages=[{"role":"user","content":[{"image":img},{"text":"Extract line items"}]}],
)

Complex reasoning about what an image implies is still weaker than Claude and GPT-5. Treat Qwen 3 VL as a perception engine; let a reasoning model draw the conclusions.