srv-sngh/Unlimited-OCR-mixed-nvfp4

Unlimited-OCR — Krill mixed-nvfp4 (MLX)

A Mac-native (Apple Silicon / MLX) mixed-precision build of baidu/Unlimited-OCR (built on DeepSeek-OCR), packaged for the Krill runtime. It parses documents and images to grounded text with no Python and no trust_remote_code — the DeepSeek-MoE language backbone and the DeepEncoder vision tower (SAM-ViT-B + CLIP-L + projector) run natively in Swift + MLX.

Run with Krill

krill pull unlimited-ocr
krill run unlimited-ocr --image your_page.png "document parsing."

It emits grounded OCR, e.g.:

<|det|>title [48, 74, 402, 130]<|/det|>Invoice 2026
<|det|>text  [33, 229, 370, 290]<|/det|>Bill to: Acme Corporation
...

Format

Mixed-precision, single-file model.safetensors (~2.3 GB, from the 6.67 GB bf16 source):

Modules	Precision
MoE experts (the residency-dominant bulk)	nvfp4 (group 16)
attention q/k/v/o, dense + shared FFN, embed, lm_head	8-bit affine (group 64)
DeepEncoder vision Linears (SAM / CLIP / projector)	8-bit affine
Conv2d kernels, norms, position/learned embeddings, router gate	unquantized

This format is Krill-specific (the native Swift+MLX runtime resolves the per-module quantization); it is not a drop-in transformers checkpoint.

Scope

Ships the base view (single 1024 global view → 273 image tokens), which reads full pages including wide layouts. The "gundam" tiling mode (local crops for very dense/large scans) is a follow-up.

Credit & license

All credit to the original authors, baidu/Unlimited-OCR and DeepSeek-OCR. MIT licensed, as upstream.