franken_ocr weights
Quantized weight artifacts for
franken_ocr
— a pure-Rust, memory-safe, CPU-only OCR engine that runs a small family of
hand-ported vision-language models with no general ML framework: no PyTorch, no
Python, no CUDA, no FFI at inference, no GPU.
These are
runtime artifacts in the project's own .focrq container, not
Hugging Face
transformers checkpoints. They are consumed by the
focr CLI,
the browser playground at
franken-ocr.com, and the
FrankenOCR iOS app. They will not load in
transformers.
Why this mirror exists
GitHub release assets return 503 under load and cap a single asset at 2 GiB,
which is why the largest artifact ships there as byte-split parts. This mirror
serves the same bytes from a CDN, with ranged requests and CORS, so a browser
can fetch a model directly and a phone can resume an interrupted download.
Every file is pinned by SHA-256 in
models/manifest-v2.json
and verified byte-for-byte before installation. A mismatched or truncated
download can never hydrate.
Files
| file | bytes | SHA-256 | recipe |
|---|
unlimited-ocr.wasm-int4.focrq | 3,003,988,117 | 2653831ccd7f481f898f80ae5c95fa1ec7ee2a5a18005d3c927ddf64ed75e187 | unlimited-ocr-wasm-experts-int4-attn-int8-lmhead-int8-v1 |
tokenizer.json | 9,979,544 | a02f8fd5228c90256bb4f6554c34a579d48f909e5beb232dc4afad870b55a8b4 | — |
unlimited-ocr.wasm-int4.focrq
Baidu Unlimited-OCR (a DeepSeek-OCR derivative: SAM-ViT-B → 16× conv token
compressor → CLIP-L/14 → linear projector → 12-layer DeepSeek-V2 MoE decoder
with Reference Sliding Window Attention) transformed from the 6.67 GB bf16
checkpoint to 3.00 GB:
- MoE routed + shared experts → int4, group size 16 for
gate_proj and
down_proj, 32 for up_proj — the measured sensitivity order.
- Attention q/k/v/o,
lm_head, embed_tokens → int8 per-channel.
- Vision tower, projector, MoE router gate, and all norms stay BF16.
Quantizing the vision encoder wrecks OCR; every downstream token is
conditioned on its 256 outputs, so error there is not local.
Quantization is calibration-aware (importance-weighted clip search plus an AWQ
down_proj fold over a 13-page activation-statistics run), measured strictly
better than plain round-to-nearest on every corpus page.
Honest accuracy note. int4 costs something real on dense material: on an
1886 newsprint page the measured character error rate is 0.156. Calibration
bought a 10.9% relative reduction on the hardest page without regressing any
page — a real improvement, not a fix. The per-page receipts live in
docs/DISCREPANCIES.md.
This artifact is not the native CLI default. focr pull installs a separate
conservative 4.16 GB artifact that keeps attention and lm_head at high
precision. This one exists for memory-constrained targets: the browser, where
the whole model must fit WebAssembly's 4 GiB linear memory, and phones.
Usage
1# CLI
2focr pull # installs the conservative native default
3FOCR_MODEL_PATH=/path/to/unlimited-ocr.wasm-int4.focrq focr ocr page.png
4focr ocr book.pdf --pages 3,5-9 -o excerpt.md
The tokenizer sidecar must sit beside the artifact.
Licenses
The engine is MIT (with an OpenAI/Anthropic rider), Copyright (c) 2026 Jeffrey
Emanuel.
The weights are a separate matter. This artifact is a quantized derivative of
Baidu's openly-licensed model, and its notice travels with it:
Baidu Unlimited-OCR — Copyright (c) 2026 Baidu, MIT License
The original checkpoint is
baidu/Unlimited-OCR
(source SHA-256
2bc48a7a110061ea58fff65d3169367eebe3aee371ca6968dc2219c1b2855fc6).
Not affiliated with Baidu
This is an independent reimplementation that runs Baidu's openly-licensed
weights. No endorsement is implied.
