franken_ocr
A pure-Rust, CPU-only, memory-safe OCR engine running a small zoo of hand-ported vision-language models, including a DeepSeek-OCR-derived model, a chart-data extractor and a sheet-music transcriber, with no Python, GPU or general ML framework required.
franken_ocr runs five models through model-specific Rust kernels: Baidu Unlimited-OCR, a DeepSeek-OCR-derived 3B mixture-of-experts vision-language model, is the fast default for document OCR, GOT-OCR2 handles specialized structured formats, SmolVLM2 handles image description and visual question answering, OneChart extracts chart data, and Polyphonic-TrOMR turns scanned sheet-music pages or staff crops into MusicXML.
All models are fetched with focr pull, which verifies each downloaded part by SHA256 before reassembly; inference itself runs offline once weights are present, needing no CUDA, FFI or GPU at inference time. The v0.8.0 release ships raw executables for macOS (Apple Silicon and Intel), Linux (x86-64 and ARM64) and Windows (x86-64 and ARM64), each with a SHA256 sidecar, and the pinned Unlimited-OCR artifact is reported to pass a 20-page corpus budget at a stated aggregate character error rate, with the project explicit that it does not yet meet its own stricter multi-signer governance bar for that claim.