← October 27, 2026 briefing · added in the 20:00 KST update
Baidu releases Unlimited-OCR 2 on Hugging Face, parsing documents with constant memory regardless of length (community report)
Announced October 27, 2026
Verified (fact)
Baidu released Unlimited-OCR 2 on October 27, according to the Hugging Face repository (baidu/Unlimited-OCR-2) description shared on Reddit's r/LocalLLaMA. The core technique is SR-SWA, a sliding-window attention method that uses sparse reference tokens. Even over very long sequences, it fits text-image attention into a fixed-size reference set, such as 512 tokens. As a result, per-step decoding latency and memory use reportedly stay nearly flat as documents grow longer. The collected items include no benchmark figures.
Needs fact-check (claim)
Baidu's claim: it parses documents of unlimited length within a finite context window.
Why it matters (interpretation)
Memory and latency growing with document length has been a bottleneck for OCR on documents hundreds of pages long. As an open model that claims to solve this at fixed cost, it is worth trying in document-processing pipelines.
Confidence medium · official source pending
Sources
- Community baidu/Unlimited-OCR-2 · Hugging Face Reddit
More about Baidu
-
Baidu says its search engine retrains itself on real queries and feedback without human tuning; prices enhanced search AI API at 4.9 yuan per 1,000 calls through year-end (report)
According to Pandaily, Baidu said its search engine retrains itself based on real search queries and user feedback without separate...
-
Baidu Health adds video/voice 'health camera' and links to 150,000 hospital doctors in AI health assistant; "50 million monthly AI users" (report)
According to Pandaily, Baidu Health has upgraded its AI health assistant. It added a camera feature that takes video and voice as input,...
-
MiniMax M3.1 Flash follow-up: Baidu-affiliated chipmaker Kunlunxin says the model runs on its accelerators, as Chinese chips race for Day 0 support (report)
According to AI Times Korea (인공지능신문), Kunlunxin, which began as Baidu's AI chip unit, said on the 23rd (local time) that it had run...