Markdown Comes to LiteParse
LiteParse 2.1 claims to be the fastest open-source, model-free PDF-to-Markdown pipeline, citing top benchmark scores across three datasets (opendataloader-bench, olmOCR-bench, ParseBench). Justy sees clear product value for teams exporting PDFs to editable Markdown. Cody questions whether these heuristics hit a ceiling and why charts/visuals are excluded.
Transcript
Justy Okay, so LiteParse 2.1 just shipped a feature that’s been requested for years — output to markdown instead of just raw text.
Cody Yeah, but it’s a heuristic tool claiming fastest open-source, model-free PDF-to-markdown pipeline…
Justy —and it’s got top scores on three benchmarks, Cody.
Cody Top among model-free tools, sure — because the benchmarks aren’t built to penalize tools that can’t recover charts or visuals.
Justy I mean, that’s the point — if you care about editability and you’re not trying to extract chart data, this is exactly what you need.
Cody Sure, if your PDF has zero charts, tables, or diagrams.
Justy Come on, most business PDFs are text plus simple tables…it’s still 80% of the problem solved.
Cody It’s the 20% that keeps biting people.
Justy Anyway — this is finally a no-model option that’s faster than every other open-source tool they benchmarked.
Cody They’re comparing apples to apples: open-source, model-free only. So fine. But that also means they disabled OCR and any model integrations when running the numbers.
Justy Of course — that’s what ‘model-free’ means. It’s all rule-based PDF parsing.
Cody Right. So the pipeline hunts for font family, font size, text position… and then runs a grid-projection algo to group lines into markdown elements.
Justy Exactly. One binary: give me the fastest way to turn PDFs into editable text without spinning up a model endpoint. This delivers.
Cody It delivers until you need charts.
Justy Which is why their benchmark tables show literally everyone scoring zero on ParseBench’s Charts column.
Cody Wait—zero? ParseBench tries to compare structured data extracted from the chart, so a heuristic engine has nothing to emit.
Justy Yeah, so they’re honest about it — those columns are basically noise. The rest of the metrics? LiteParse leads overall.
Cody Mm-hm. 0.328 overall on ParseBench vs the next best at 0.310. Tiny margins.
Justy Tiny but repeatable across three datasets. opendataloader-bench at 0.875, olmOCR-bench at 0.391.
Cody Big spread on olmOCR — that’s mostly optical-character-recognition accuracy, not layout or markdown fidelity.
Justy Still, they’re measuring what matters for day-to-day use: content faithfulness, semantic formatting.
Cody Content faithfulness jumps out — LiteParse at 0.686 versus pymupdf4llm at 0.609. That’s meaningful.
Justy And for teams that just need the text out of PDFs so someone can edit it, this is the answer.
Cody If there are no tables or charts.
Justy I labored over a client deck all Thursday because the exported text lost the table layout. This would’ve saved me half the headache.
Cody …while I was up at two a.m. last night fighting GLM-5.2 docs and a four-hour sleep budget.
Justy Yeah. Implementing LiteParse’s markdown mode in your export pipeline would’ve let you skip the whole font-mismatch nightmare.
Cody Okay okay. Install’s trivial: pip install liteparse, then lit parse doc.pdf --format markdown.
Justy Or the Python helper: LiteParse(output_format='markdown').parse('doc.pdf').
Cody And they even give you a demo running in-browser with WASM.
Justy Real product value lands for anyone drowning in PDFs they need to edit — legal, contracts, research PDFs, invoices…
Cody …until someone opens a PDF with a chart and asks why it’s missing.
Justy Then you tell them: ‘this is model-free,’ and they go spin up a model parser. Meanwhile, eighty percent of the day-to-day docs are handled without a billable model.
Cody Fair.
Justy I’m downloading it now. You in?