do you think frontier multimodal models will eventually replace the dedicated
retained text
honest question not a hot takeway most extraction pipelines still run a dedicated parser or ocr step like docling, marker, llamaparse, Surya whatever to turn the pdf into clean text or md before feeding llm but frontier multimodal models are getting more enhanced at just reading the page directly and cheaper every release so this seems worth asking where this actually goes case for the models absorbing it: capability keeps climbing and native pdf/vision handlng is built in now +dropping the parser is one less movng part and one less dependency to babysit . Case for the parser layer sticking around - layout, tables and grounding like which cell a value belong to and pointing back to the exact region on the page are a different problem than reading text and the benchmarks on that are still contested Not tryna push either side just wanted to compare my thoughts vs others, how do you guys see it going
document fetched 2026-09-24 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wo1ott/do_you_think_frontier_multimodal_models_will