OCR Is Research Again

🧩Three Technical Bets

Same goal, divergent architectures

Mistral OCR 4, PaddleOCR PP-OCRv6, and Baidu Unlimited-OCR aim at the same outcome but pick different levers: a hosted multilingual model, a sub-35M-param on-device family, and a constant-KV-cache long-horizon parser. The divergence is the story.

🌐Mistral OCR 4

Hosted, multilingual, structure-aware

Mistral OCR 4 leans on broad language coverage and document-structure output as a hosted API, positioning OCR as a priced service rather than a model you run. It competes on convenience and table/layout fidelity over raw character accuracy.

open_in_new mistral.ai/news/ocr-4

📦PP-OCRv6

50 languages under 35M params

PaddleOCR's PP-OCRv6 ships a tiered family small enough to run on-device while still covering 50 languages. It is the counter-thesis to hosted VLMs: detection and recognition as a tiny, embeddable component, not a remote call.

open_in_new huggingface.co/blog/PaddlePaddle/pp-ocrv6

📜Unlimited-OCR

One-shot, page-uncut parsing

Baidu's Unlimited-OCR keeps a constant KV-cache to parse long documents in one pass instead of slicing pages. It reframes OCR as a long-horizon decoding problem, where context length and memory shape what a single forward pass can read.

open_in_new github.com/baidu/Unlimited-OCR

🔬Engineering Deep-Dive

The convergence written up

A research write-up tracing how these three releases reframe document OCR — small specialists vs general VLMs, page-uncut long-horizon parsing, and what to actually measure once benchmarks saturate.

⚖️Specialist vs General VLM

Where the tradeoff actually bites

A document-only model trades breadth for cost and latency; a frontier VLM reads anything but pays per token and drifts on tables. The choice is rarely about peak accuracy — it's about cost per thousand pages and how the output plugs into downstream parsing.

💸Cost Per Page

The real selection axis

General VLMs charge per token over the whole rendered page; specialists charge a flat tiny cost. At document scale the gap is several-fold, which is why a worse-on-paper specialist often wins the production slot.

🧾Structure Fidelity

Tables, math, reading order

Plain character accuracy hides the hard part: reconstructing tables, equations, and reading order. Models that emit structured layout, not just text, are the ones that survive contact with real PDFs and scanned forms.

🛠️Pipeline Placement

What goes where in a RAG/doc stack

OCR sits upstream of chunking, embedding, and retrieval. Picking the wrong model there propagates errors everywhere downstream, so placement decisions — pre-extract vs VLM-in-the-loop — matter more than a single benchmark point.

🔎Pre-Extract vs In-Loop

Two integration patterns

Either OCR runs once up front and feeds clean text into the index, or a VLM reads the page on demand inside the agent loop. Pre-extract is cheaper and cacheable; in-loop is flexible but pays VLM cost on every query.

🧵Error Propagation

Why upstream choices compound

A mis-read table at OCR time becomes a wrong embedding, a bad retrieval, and a confidently wrong answer. The cheapest place to fix document quality is at extraction, not in prompt patches later.

📊The Benchmark Ceiling

OmniDocBench is saturating

When several models cluster above 90 on the same suite, the leaderboard stops being a signal. The interesting differences move to multilingual coverage, structure fidelity on tables and math, and dollars per page — none of which a single saturated score captures.

🗣️Multilingual Coverage

The next real differentiator

As English-document scores converge, language breadth becomes the separator. A model strong on Latin scripts but weak on CJK, Arabic, or Indic text is a different product than one trained for global coverage.

❓What To Measure Next

Beyond a single score

The open question is what replaces a saturated leaderboard: per-language accuracy bands, structure-level F1, and cost-normalized quality. Until those are standard, model choice stays an engineering judgment, not a ranking lookup.