We built the fastest PDF -> markdown parser in the world 🚀⚡️ AND it’s more accurate than any other open-source, model-free parser (pymupdf4llm, opendataloader, pdf-inspector, markitdown) on 3 standardized benchmarks: olmOCR0-bench, opendataloader-bench, ParseBench Introducing LiteParse v2.1. The v2 base version was already the fastest document->text parser on the planet,... and with this new release we’ve introduced markdown. It is fully open-source (Apache 2.0) and free, is usable from CLI/Rust/Node/Python/WASM, and is also installable as a one-click agent skill. Check it out: Come check out LiteParse:show more

Here are our accuracy benchmarks

Here are our speed tests

How are you handling images?

This is great. I tested this against Markitdown for a 2-column scientific paper (ACM format) with figures/images. Both have OCR capabilities, but Markitdown is basically unreadable compared to LiteParse. That said, there's still a lot of weirdness to fix on output structure.

awesome to hear. we'll keep iterating on the capabilities going forward 🫡

PDF转markdown这块的坑我们做数据的太有体会了 表格和多栏一崩 后面整个pipeline都得返工。最快只是入场券 真正难的是它在脏扫描件上还稳不稳 那才是日常会要命的地方

how do you’ll deal with very visual PDFs (drawings/blueprints etc)

Fastest, most accurate and oss free! How can anyone beat that combination -:) Great to see continued innovation!

Added to my app CIP🔥🔥

Fast on clean PDFs. The harder problem: scans with bad lighting, folded pages, handwriting mixed in. That's where most parsers fail.

Fast, model-free parsing is exactly what agents need at the tool layer — every LLM call you save on doc prep is latency and cost back in the budget. Clean inputs beat a bigger model more often than people admit.

cc @MatthieuTalbot pour paranoia

open source tools are best

The interesting shift is that AI workflows are becoming more like software pipelines. Better parsing improves the input layer, but the experiments built on top of that need the same level of organization and iteration.

@llama_index Can we reverse engineer an image or a PDF, even if it is rasterized, back to TeX in one shot? That, IMHO, would be the end game for OCR/VLM.

Всем плевать на скорость. Главное это качесивл

Document parsing is the unglamorous wall every real-world AI pipeline hits. We process real estate contracts and title docs, and a bad parse quietly poisons everything downstream. Model-free and fast matters more than people admit at scale. Testing this on our stack.

This is great, I was looking for an alternative as pymupdf 4llm was taking too much memory Will need to test it on my data though, how well does it perform on table heavy pdfs? Does it preserve them?

it should preserve the overall text layout as of v2, and markdown structure as of v2.1. the markdown piece is new though so please report if any feedback/issues!

Its been day 2 and I am noticing something weird. Pros so far: My image size went down, Liteparse is definitely fast Biggest con: It is not as stable as pymupdf, as for shorted page windows it looses headers. I have been trying to test liteparse on the small machine I was running pymu and I am seeing much more failures. So am stuck optimising batch size Vs ram in my VM at the moment.

cc @LoganMarkewich to help take a look. If you open a github issue we would be happy to look at in order to help improve

@LoganMarkewich Here you go:

How do you deal with a pdf that has an image inside with text? Is it possible to detect this before passing through this tool, extract image and send to vision model?!

Parsing speed matters when documents become live inputs.

open source tools like this save so much time

You ran three parser benchmarks. Would you sanity-check one agent benchmark for me? The big drop comes from fewer agent rounds, not trimming output:

How does it compare to Docling please 🙏

Have you ever heard intrusive thoughts?

@llama_index Would love to see how it compares to Mistral OCR 🧐

Nicely explained. The section around Markdown being easier to inspect, copy, diff, and reuse is exactly why this kind of workflow matters for teams. The official MarkItDown Online website is here for reference:

bun build --compile のサポートを追加できませんか？ ネイティブモジュールへの依存が大きく、適切な統合がされていないためBunのバンドラーがネイティブモジュールを認識できずバンドルされません

@jerryjliu0 A sharper test for "We built the fastest PDF -> markdown parser in the world 🚀⚡️ AND it’s more accurate than any other open-source, mo...": show one visible before/after, one failure case, and one repeatable check. What would make this credible enough to use?

support scanned pdf?

PDF parsing is one of those problems people only notice when it break

liteparse v2.1 claims to be faster than others but doesn't mention how much faster, we'd need to benchmark against actual outputs, not just claim it's "fast

can you start a 💬?

markdown tables for logs kept breaking during parsing. does v2.1 handle those edge cases?

the "no LLM" part is what actually sells me. model-based parsers always look amazing in a demo then fall apart on cost the moment you run them at real volume. fast + deterministic + accurate is the boring combo everyone quietly needs

wow , Jerry Liu I was facing issue with markdown I want to share docs in form of pdf Its really super fast

Check please your dm