Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing

The Avocado Pit (TL;DR)

• 🚀 Baidu's Unlimited OCR is a 3B-parameter model that parses long documents like a boss.
• 🧠 Its R-SWA tech keeps memory use and latency steady no matter the output.
• 🏆 Scored 93.23 on OmniDocBench, crushing the previous record by 6.22 points.

Why It Matters

Baidu just dropped a tech bombshell with their Unlimited OCR, and it's not just a bunch of zeros and ones. This isn’t your average OCR model—it’s like having a superpower that lets you read War and Peace in a single glance without breaking a sweat. How? By keeping the Key-Value (KV) cache flatter than a pancake on a lazy Sunday morning. This means more efficiency and speed, less drama.

What This Means for You

Whether you're scanning legal documents, analyzing financial reports, or just don’t want to waste time waiting for your computer to catch up, Baidu's Unlimited OCR has your back. Its capability to handle long documents without the usual memory bloat is a game-changer for businesses and tech enthusiasts alike. It’s like having a personal document wizard on your team.

The Source Code (Summary)

Baidu’s Unlimited OCR is an open-source marvel packing 3 billion parameters into a Mixture of Experts (MoE) model. It uses Reference Sliding Window Attention (R-SWA) to keep the KV cache flat, ensuring that memory and latency don’t skyrocket with longer documents. This innovation scored a cool 93.23 on the OmniDocBench v1.5, leaving the previous DeepSeek OCR baseline in the dust by a whopping 6.22 points. Under an MIT license, this tech is as free as a bird, inviting anyone to give it a whirl.

Fresh Take

In a world where digital clutter is the norm, Baidu’s Unlimited OCR comes as a breath of fresh air. By tackling the classic OCR bottleneck—memory and speed—head-on, it sets a new standard. While many models choke on long documents, Baidu's latest brainchild says, "Bring it on!" This could lead to more seamless integration of AI in industries that rely on heavy document processing. So, if you're in the document-heavy trenches of work life, Baidu's got the artillery you need to come out on top. Plus, it’s open source, so it's like getting a golden ticket to the chocolate factory—if the factory made really smart robots instead of candy.

Read the full MarkTechPost article → Click here