Breaking

Baidu Open-Sources Unlimited-OCR, a 3B Model That Parses Entire 100-Page PDFs in a Single Pass

July 20, 2026 at 02:15 EDT

Foundation Models
Open Source
Media Generation

Baidu has released Unlimited-OCR, an MIT-licensed 3-billion-parameter document model that can read an entire multi-page PDF in one forward pass on local hardware, sidestepping the page-by-page chunking that has long constrained end-to-end OCR. The model, published on June 22, 2026, targets a stubborn bottleneck in machine reading: how to process long documents without runaway memory use or degraded accuracy as output length grows.

June 22, 2026 · Baidu

Unlimited-OCR Reads a Whole PDF in One Pass

An MIT-licensed, 3-billion-parameter document model uses constant-size memory to parse multi-page PDFs end-to-end on a single GPU — no page-by-page chunking.

total parameters — only ~500M active per step (MoE)

32,768

token context window for multi-page inputs

1 GPU

runs locally with memory bounded regardless of output length

OmniDocBench accuracy vs. baseline

87.2%

DeepSeek-OCR
(baseline)

93.23%

Unlimited-OCR
(v1.5)

93.92%

Unlimited-OCR
(v1.6)

~6 points above baseline · +12.7% throughput · error rate below 0.11 on 40+ page documents (vendor-reported)

The mechanism · R-SWA

Reference Sliding Window Attention keeps the KV cache at a constant size — so memory no longer grows with output length.

Visual reference tokens
use full attention

Generated text
attends to a fixed sliding window

→

Constant KV cache
one-shot long-horizon parsing

Where it works

Solid on printed text, tables and equations
Local alternative to metered cloud OCR — no per-page charges
MIT license fueling fast community uptake

The caveats

Headline accuracy is vendor-reported, awaiting outside confirmation
Degraded accuracy and hallucinations on handwriting
Specialized systems can still beat it on narrow tasks

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.