Cathy Fu

Global Comms Lead @Baidu_Inc, prev. @BytedanceTalk; alum of @PKUYCA and @DBA_Programme
The usual disclaimers

[Cathy Fu @cathyf30]

@CHRobotaxiWatch The RTA designates multiple zones for AV deployment in Dubai. Zone 1 covers Jumeirah, Umm Suqeim and Al Wasl areas while Zone 2 & 3 includes parts of Downtown Dubai, Business Bay & more. Lots of exciting updates to come this year for sure : )

[Rob Grant @RobGsix]

@Baidu_Inc Great transparency! Explains why you rank #1 on the @RoadToAutonomy Robotaxi Index

https://t.co/SvSbFyqY9Q

[Baidu Inc. @Baidu_Inc]

What's new with Apollo Go? Swiss roads, a wider Dubai footprint, and more kilometers logged in Hong Kong.

The full rundown 👇

[Elias @iam_elias1]

Baidu just open-sourced an OCR model that reads entire books in one go.

3 billion parameters. 500 million active. Runs on a single GPU. MIT licensed. Free.

It is called Unlimited-OCR. It launched June 22, 2026. 1,800 GitHub stars in the first 24 hours. And it solves the most frustrating problem in document AI.

Here is the problem every developer building document pipelines hits.

Most OCR models read one page at a time. You split a PDF into individual images. Run each page through the model separately. Then stitch the results back together and hope the seams do not break anything.

Tables that span page breaks get mangled. Cross-page references get lost. Headers and footers contaminate the output. The stitching code becomes its own maintenance burden. And every page resets the model's memory so it has no idea what it read on page 3 when it is processing page 4.

This is how every OCR pipeline in the world has worked for years. Chunk, process, stitch, pray.

Baidu just eliminated the entire loop.

Unlimited-OCR ingests up to 40 pages in a single forward pass. The model sees the entire document at once. Tables that span pages are handled correctly. Cross-page context is preserved. No chunking. No stitching. One pass in, structured text out.

Here is the technical innovation that makes this possible.

Traditional OCR models use standard transformer attention which means every new token the model generates looks back at everything it has already generated. For a short document, that is fine. For a 40-page PDF, the memory required to store all those lookback keys and values, the KV cache grows so large that the GPU runs out of memory or slows to a crawl.

Baidu's insight came from watching how humans actually copy long documents. When you transcribe a book, you do not re-read every word you have already written before writing the next one. You remember a window. Older content fades. You keep writing.

They built a new attention mechanism called Reference Sliding Window Attention- R-SWA - that does the same thing. The decoder retains full access to the compressed document images but only attends to a fixed sliding window of recently generated text. The KV cache stays constant. A 40-page document costs the same memory as a 2-page one.

The model processes page 40 just as efficiently as page 1.

Here is what the model actually outputs.

Not raw text. Structured, layout-aware output. Tables come out as HTML. Equations come out as LaTeX. Reading order is preserved. The output respects the document's actual structure- columns, headers, footers, captions instead of dumping everything into a flat text stream.

For anyone building RAG pipelines, this changes the quality of what enters your retrieval layer. Better parsing means better chunks. Better chunks mean better retrieval. Better retrieval means better answers.

Here is how the benchmarks look.

93.23 on OmniDocBench v1.5, beating the DeepSeek OCR baseline it was built on by 6.22 points. Edit distance stays below 0.11 even at 40+ pages. 96.90% text diversity meaning the model is not collapsing into repetitive output on long documents.

Unlimited-OCR was built on top of DeepSeek OCR via continue-training approximately 2 million document samples across 4,000 training steps. Not a from-scratch run. A targeted improvement on an already strong foundation.

Convert your PDF pages to images with PyMuPDF first, then feed them all in at once. One call. Every page. Structured output.
It also runs on vLLM, SGLang, and Docker Model Runner. Weights are on Hugging Face and ModelScope.
Here is the honest part most posts leave out.

40 pages is a soft ceiling, the 32K context window caps it. Handwriting and non-standard fonts remain weak. Scientific PDFs converted to LaTeX and electronics datasheets with merged table cells are still hard. One OCR product founder with a decade of experience put it bluntly on Hacker News: "OCR still sucks in 2026."
He is right about the edge cases. But for contracts, invoices, reports, research papers, legal filings, and every standard business document that enters an AI pipeline, Unlimited-OCR just made the parsing step dramatically simpler, faster, and free.
Here is the comparison that matters.
Mistral OCR 4 launched the next day, June 23. It wins on raw accuracy and supports 170 languages. But it costs $4 per 1,000 pages and requires a cloud API. Your documents leave your machine.
Unlimited-OCR is free. Runs locally. Your documents never leave your device. For anyone processing sensitive contracts, medical records, financial documents, or legal filings that is not a feature. It is a requirement.

4,663 GitHub stars. MIT licensed. One GPU. 40 pages in one pass. Free.
The era of page-by-page OCR just ended.

Source: Baidu · MarkTechPost · ExplainX · ByteIota · AlphaMatch · June 22, 2026
( Link in the comments)

[Bloomberg @business]

In the China tech space, Baidu is now a full-stack player in the AI industry. As the maker of its own chips, AI model (Ernie), cloud system and a self-driving car business (Apollo Go), the tech giant is a very desirable company to work for.

On this episode of the Odd Lots podcast, Baidu CFO Henry He joins @TheStalwart and @tracyalloway to discuss how the company uses AI to evolve its own organizational structure and attract new talent. Listen at https://t.co/RqFaugaTL4 or watch at https://t.co/fcjUvfje6d.

[Cathy Fu @cathyf30]

Thanks for riding with us! This coastal route looks perfect🏖️

[China Robotaxi Watch @CHRobotaxiWatch]

Just took a @Baidu_Inc Apollo Go ride in Jumeirah, Dubai.
Fully driverless rides across the whole district are showing up for 5 AED. About $1.36.

No safety driver, major tourist hub, less than the price of a coffee.

Pretty wild.🤯

[Baidu Inc. @Baidu_Inc]

https://t.co/yWwQ1jOMox

[Bloomberg @business]

From Internet search to chips, LLMs, clouds, and self-driving cars. https://t.co/I1JYfPf5BB

[Bloomberg @business]

In the China tech space, Baidu is now a full-stack player in the AI industry. As the maker of its own chips, AI model (Ernie), cloud system and a self-driving car business (Apollo Go), the tech giant is a very desirable company to work for.

On this episode of the Odd Lots podcast, Baidu CFO Henry He joins @TheStalwart and @tracyalloway to discuss how the company uses AI to evolve its own organizational structure and attract new talent. Listen at https://t.co/RqFaugaTL4 or watch at https://t.co/fcjUvfje6d.

[Baidu Inc. @Baidu_Inc]

That's a wrap on the 2026 Hong Kong Auto Expo!

As Hong Kong's first autonomous vehicle pilot licence holder, Apollo Go has logged more than 230,000 km of safe testing in the city — its first right-hand-drive market — as of April 2026. Over the past year, our autonomous vehicles have expanded from Airport Island to North Lantau, Kowloon East, and Southern Hong Kong Island.

[Cathy Fu @cathyf30]

@gbrulte @Baidu_Inc Approximately 25 km² now. More updates coming soon!

[TechCrunch @TechCrunch]

TechCrunch Mobility: A new robotaxi scorecard shows China’s dominance https://t.co/HlAKoZaE3E

[The Road to Autonomy Indices @AutonomyIndices]

The Road to Autonomy Robotaxi Index update for the week of June 19, 2026.

https://t.co/ixZB3hFq6t

[Cathy Fu @cathyf30]

@gbrulte @Baidu_Inc Approximately 25 km² now. More updates coming soon!

[Cathy Fu @cathyf30]

@gbrulte @Baidu_Inc Fully driverless now without safety operators : )

[Joe Weisenthal @TheStalwart]

Just recorded our first ever live Odd Lots in Asia with Baidu CFO Henry He

We talked a lot of AI (obviously) but also, as I wrote in the liveblog, the global self driving car race Is going get super interesting as Baidu and Waymo compete in the same cities.

[Baidu Inc. @Baidu_Inc]

Zack Williams brought one of the world's oldest writing systems to the ERNIE AI Developer Challenge: ancient cuneiform tablets.

Using PaddleOCR, he built NabuOCR to help read cuneiform from tablet images.

See the story behind @BoatbomberRBLX's winning project ���

[Poe Zhao @poezhao0605]

A small but useful update from Baidu Create 2026:

Baidu says its Kunlunxin P800 chips have now been validated at scale and used in multiple 10,000 card clusters since 2025.

It also says a Kunlunxin based cluster trained a key ERNIE 5.1 version, reaching 97% effective training efficiency and over 85% linear scaling at 10,000 card scale.

Baidu’s 256 card Tianchi supernode is scheduled for June. The company says it has been adapted for ERNIE, DeepSeek, GLM, MiniMax and other major models.

[Cathy Fu @cathyf30]

@heyshrutimishra Thank you for coming!

[Shruti @heyshrutimishra]

At Baidu Create today in Beijing. 🇨🇳

Robin Li's keynote made the pattern clear: Baidu chip, Baidu cloud, Baidu model, Baidu app.

A vertical AI company, every layer built inside China.

What stood out:
• Kunlun chips: 30,000+ cards in production, Kunlunxin IPO incoming
• ERNIE 5.1: pre-training cost cut 94%
• Agents at scale: DuMate, Miaoda, Huiboxing
• "Disposable software": built for one task, used once, thrown away
• 10B daily active agents forecast, 3x Meta's DAU

This wasn't any single launch. It was the whole stack, in one company, in one country.