Qunyi Xie

Work @ Baidu OCR Team

[Qunyi Xie @XieQunyi]

HF: https://t.co/Z3ZRriItcK
Git: https://t.co/lMm6gJyy9w

[Qunyi Xie @XieQunyi]

Meet HPD-Parsing: a new high-throughput document parser built on Hierarchical Parallel Decoding.

⚡️The speed numbers speak for themselves — we hit a max 4,752 TPS throughput! That’s 2.62x the fastest current top document parser, and a whopping 3.06x faster than our AR baseline.

[Qunyi Xie @XieQunyi]

It's realy a OCR moment！

[Youyang Yin @HYPERUU4122]

Thanks @ylecun for reposting our work.

I stared at the notification for a few seconds before it really sank in. Still processing it.
But this feels like a good moment to share what Unlimited-OCR was really about.
Most people assume OCR is a solved problem: detect text, recognize text, process page by page, done.

But we kept coming back to one simple question:
How do humans copy long documents?
Not by replaying every previous word in memory. You glance at what you just wrote, recover your place, and continue.
That’s how people handle long sequences.
OCR systems usually do the opposite. They process one page, reset everything, and start over on the next. Page after page.

That isn’t intelligence. It’s a for loop with amnesia.
So we built R-SWA: an attention mechanism that preserves full access to the source image, while keeping only a short sliding window over the generated text.
The intuition is simple:
A model does not always need to remember everything. Sometimes it just needs to know where it is.
That small shift gave us something surprising:
KV cache no longer grows with output length
long-document inference remains stable
performance stays strong
For this task, restricting attention was not a tradeoff. It was a free lunch.
And maybe this idea matters beyond OCR.

How many long-context problems are really about position, rather than perfect recall?
ASR, translation, and more.

A deeper lesson here may be this:
The future is not necessarily models with infinite memory. It may be models that learn how to forget—elegantly.

Also, huge thanks to @VaibhavSisinty for the thoughtful summary.

[Qunyi Xie @XieQunyi]

Public interest in the story continues to grow.👉

[Youyang Yin @HYPERUU4122]

Unlimited-OCR has surpassed 1M downloads on Hugging Face 🚀

❤️ Huge thanks to everyone supporting the project
👀 Next up: multilingual and 128k context versions

GitHub: https://t.co/VmU9Fhy6A3
HF: https://t.co/hRIm2kzsBS

[Qunyi Xie @XieQunyi]

Unlimited-OCR sparked a wave of end-to-end doc parsers — OCR is now the boilerplate coding grind for VLMs.

[Qunyi Xie @XieQunyi]

Humans, can easily parallelize transcription through multi-person collaboration, only needing agreement on segment assignments beforehand.

How about a VLM based document parser？

[Qunyi Xie @XieQunyi]

Good News👏Follow Talk for more updates.

[Youyang Yin @HYPERUU4122]

Excited to see vLLM v0.25.0 out! 🎉

🚀 Thrilled that Unlimited-OCR has landed in the official vLLM release.

We’ll be sharing more about Unlimited-OCR at WAIC:

📍 Baidu booth, Hall H1, 1F, Shanghai World Expo Exhibition & Convention Center
🕦 July 18, 11:30 AM–12:10 PM (BJT)

[Qunyi Xie @XieQunyi]

For LLM based document parser, inference efficiency is severely constrained due to auto-regressive decoding mechanism.

The two key points are attention length and sequential waiting. The former is well addressed by Unlimited-OCR. While P-MTP significantly mitigates the latter.

[Qunyi Xie @XieQunyi]

Unlimited-OCR is now officially launched on the Baidu Intelligent Cloud official website: https://t.co/7LhtqZ08Jw

[Youyang Yin @HYPERUU4122]

Unlimited-OCR has surpassed 1M downloads on Hugging Face 🚀

❤️ Huge thanks to everyone supporting the project
👀 Next up: multilingual and 128k context versions

GitHub: https://t.co/VmU9Fhy6A3
HF: https://t.co/hRIm2kzsBS

[Qunyi Xie @XieQunyi]

Congratulations! Stay tuned for ongoing updates on Baidu's OCR technologies.

[Youyang Yin @HYPERUU4122]

Unlimited-OCR has surpassed 1M downloads on Hugging Face 🚀

❤️ Huge thanks to everyone supporting the project
👀 Next up: multilingual and 128k context versions

GitHub: https://t.co/VmU9Fhy6A3
HF: https://t.co/hRIm2kzsBS

[Qunyi Xie @XieQunyi]

Our work P-MTP has been accepted by ECCV 2026. Stay tuned!

The first successful validation of MTP depth scaling of document parsing, achieves a speedup up to \(5.24\times\).

Paper: https://t.co/scmhKg5LvE
Code：https://t.co/8X70CRCoP4