I Ran a 3B Unlimited OCR Model FULLY LOCAL (Ollama Couldn't) — DeepSeek-OCR on Windows

A brand-new 3-billion-parameter OCR model that turns documents, tables, and even handwriting into clean markdown with bounding boxes — running fully local on a laptop GPU. The catch: Ollama can't load it yet. So this video is the real, unedited build from source on Windows, including every wall I hit and the exact fix.

Model: DeepSeek-OCR, packaged as the "Unlimited-OCR" GGUF — huggingface.co/baidu/Unlimited-OCR.

Why bother

OCR APIs charge per page and see your documents. A local 3B OCR model on an 8 GB laptop GPU means unlimited pages, private by default, at ~300 tok/s. And this one doesn't just read text — it reconstructs structure: tables come out as HTML, layouts keep their reading order, and every element gets a bounding box.

The build (and the three walls)

Ollama refuses the architecture, so it's llama.cpp built from source with PR 17400. Three walls on the way, all on camera:

None of these are exotic — they're exactly the walls you'll hit on a stock Windows 11 machine, which is why I left them in.

Tested on: Windows 11, RTX 4060 Laptop (8 GB), CUDA 12.9, ~300 tok/s.

The demos

Four real documents, no cherry-picking:

• Invoice — fields, amounts, layout → clean markdown
• Table → HTML — an actual <table> with rows and columns intact
• Handwriting — legible cursive comes out as text
• Multi-column paper — reading order preserved across columns

Honest cons

It's a 3B model, not magic: very dense or degraded scans can trip it, and the toolchain is bleeding-edge — you're building a PR branch, not installing an app. When Ollama support lands this becomes a one-liner; today it's a 20-minute build.

Verdict

Local OCR at this quality was a paid-API-only capability six months ago. If you process invoices, papers, or scans and want them as structured markdown without a per-page bill, this is worth the build friction. Every command and fix is in the video.