3B-Param OCR Model Runs Entirely in Your Browser via Rust & WASM
doodlestein · x · 2026-08-12
A developer introduced frankenocr, a 3-billion-parameter OCR project that runs entirely locally within a browser tab.
- Core: Hand-ports vision-language models like Baidu's Unlimited-OCR into memory-safe Rust, compiled to WebAssembly for pure CPU inference. No PyTorch, Python, or GPU required.
- Privacy: Document processing is 100% client-side with 0 bytes uploaded, enforced by browser CSP.
- Models: Includes a main model for Markdown conversion (2.8 GB int4), sheet music recognition (58 MB), formulas/tables (776 MB), and photo description (1.0 GB).
- Trade-off: Processing a single page takes 80–250 seconds on desktop, prioritizing extreme privacy over speed.
Related event: FrankenOCR Enables High-Accuracy Text Recognition in Browsers(2 posts)→
More from coding & agent
- Indie Dev Warns: CTO's Bias Against AI Coding Cost Them 6 Months — daluoseo · 2026-08-12
- ChatGPT Prompts + Local H3: Creating a 113s Short Film Entirely Offline — CeFurkan · 2026-08-12
- Local Model Drives Multi-Agent Collaboration: Inside the Clairvoyance Project — draginol · 2026-08-12
- Train a Coding Agent in One Command: HF Unveils Remote Sandbox RL Solution — SergioPaniego · 2026-08-12
- Why Agents Fail Silently in Production: The Hidden Danger of Sanitized Memory — Alternative_Pin9598 · 2026-08-12
- Build a Life OS with a Single Prompt: Hermes Agent Architecture — Teknium · 2026-08-12