@berryxia: Wow, this move directly poached DeepSeek's talent! Last night I saw this interesting OCR open-source model on HuggingFace and the fascinating story behind it. This OCR model is completely different from traditional ones! Its speed and accuracy are absolutely unbeatable~~ Let me start with some background, for those who are familiar…

Whoa, this move directly poaches from DeepSeek’s “talent pool”!

Last night, I saw HuggingFace highlighting this interesting open-source OCR model along with the fascinating story behind it.

This OCR model is completely different from traditional OCR models! Its speed and accuracy are simply unbeatable.

First, some background. As regular readers know, I’ve run several OCR evaluations recently (check my earlier posts), testing 18 documents across 6 scenarios and building local workflows. So I have some feel for OCR’s capabilities.

The biggest headache in those evaluations wasn’t accuracy — it was the workflow for multi-page documents. Every model processes pages one by one.

Each page clears memory, then an external scheduler stitches the results together. It’s essentially a for-loop, not true long-range understanding.

But Baidu’s newly open-sourced Unlimited OCR takes a completely different approach. It doesn’t process page by page.

With a single forward pass, dozens of pages are transcribed in one go.

The core value proposition is simple: One-Shot Long-Horizon Parsing. In other words, it enables deep syntactic understanding of long texts without massive labeled data, at low cost, and adapts to few-shot capabilities of large language models.

You can just throw in an image or a multi-page PDF and parse it all at once — no need to chop it into pieces and run repeatedly. This is genuinely awesome!

Apparently, the model’s inspiration is interesting. When humans copy a book, they don’t memorize the entire book.

They focus on only three things: the original text, the few words just written, and the next words to write. Earlier content naturally fades out. Recent context tracks progress. This everyday behavior reveals an attention pattern very different from current models.

Unlimited OCR’s core mechanism, R-SWA (Reference Sliding Window Attention), simulates this process.

Every token sees the full image, but the output side only maintains the previous 128 states. With 32K context, it processes dozens of pages in a single inference. The KV Cache size remains constant — it doesn’t grow with document length.

This essentially pushes OCR from a character recognition tool toward a document understanding engine. People used to think long documents must be chunked.

Now it’s becoming clear: as long as context is long enough and the model is strong enough, going end-to-end in one shot is more efficient and accurate.

The technical report is also written in a very engaging, story-driven, and bold style. It has an explorer’s vibe — previously a hallmark of DeepSeek’s technical reports.

And then things get interesting.

I looked up the key contributors of the technical report. Of three people, two use real names. But the technical director goes by a two-letter abbreviation: YY. Who is YY?

Let’s trace back.

The GitHub acknowledgments list DeepSeek-OCR and DeepSeek-OCR-2 at the top two spots. DeepEncoder was originally introduced in DeepSeek OCR.

This Unlimited OCR happens to perfectly integrate that high-compression encoder.

The mentions of DeepSeek OCR in the report don’t sound like benchmarking against a competitor. They sound more like reflecting on and improving their own previous research.

The OCR community in China is not that large. The number of people who could achieve a breakthrough like R-SWA and also have hands-on familiarity with DeepSeek OCR’s architecture can be counted on one hand.

Let’s look at another detail.

On April 24, 2026, DeepSeek-V4 was officially released. At the end of the 58-page technical report, nearly 300 names are listed in alphabetical order.

Ten of those names have a small asterisk next to them: “Resigned.” From the second half of 2025 to early 2026 — less than half a year — five people left DeepSeek.

Where did they go? Who is YY? The report doesn’t say directly, but the more you read, the more the answer seems to lie between the lines.

It’s also clear that Baidu’s recent path is indeed different. You must know that their OCR has always been the strongest — almost no competition!

From PaddleOCR to this Unlimited OCR, you can feel them moving in a more cutting-edge direction.

This update speed, this talent pipeline, and the development direction — the future looks promising.

Putting the gossip aside and focusing on the tech: the end-to-end long-document OCR direction is definitely right.

The project and model are open-source. If interested, check the links in the comments.

Berryxia.AI (@berryxia): This speed is insane! Whoa!

The newly open-sourced Unlimited-OCR can process hundreds of pages in one shot, and the speed remains stable.

This model comes from Baidu’s recent release on Hugging Face. Its core innovation is R-SWA (Reference Sliding Window Attention).

It keeps the KV Cache constant during decoding, preventing explosive growth as the number of document pages increases.