Indra OCR is a lightweight (≈1M-parameter) OCR package optimized for extremely high CPU throughput and full-page layout extraction. On 32 x86 CPU cores it processes 49.4 pages/second and gives single-page latency around 230 ms; on a per-page benchmark (OmniParseBench) its default mode scores headline 62.1, text 41.3, tables 56.4 and layout 88.6 at a median 228 ms/page. It returns line and block bounding boxes, block labels (headers, lists, tables, captions, footnotes, forms, etc.), reading order, and clean markdown or JSON. When a PDF has a text layer it uses that instead of OCR (median 130 ms for text-layer pages). An optional math mode boosts text and headline scores (text to 57.5, headline to 66.8) at the cost of higher latency (~360 ms).
Compared with other engines, Indra OCR runs 11-17x faster than Tesseract and PaddleOCR PP-OCR text models and far faster than PaddleOCR’s full PP-StructureV3 pipeline while matching or exceeding its headline, table and layout scores. Throughput scales roughly linearly with cores (7.3/14.2/27.2/49.4 pages/sec on 4/8/16/32 cores); a 300-page job including model loading runs at ~37.7 pps on 32 cores. Tradeoffs include weaker handling of complex multi-level tables, poor handwriting recognition (~8% on handwriting tests), limited script coverage on scanned images (Latin, Cyrillic, Greek, Devanagari, Bengali and Chinese only), no vertical-text support, and math off by default. Code and models are available on GitHub and Hugging Face.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
AlphaNIST is an open-source project aimed at creating a standardized dataset for evaluating and benchmarking AI models. It is designed to facilitate research and development in artificial intelligence by providing a common testing framework. (alpha-nist.github.io)
PostEra is hiring AI researchers to develop agentic systems that automate drug discovery processes using machine learning. The role involves building models to analyze biological data and support decision-making in medicinal chemistry and biology. (jobs.lever.co)
Parsimony is an open-source benchmark that measures how much code AI models add, penalizing excess code to promote cleaner solutions. It ranks models based on net code units added, encouraging efficient coding practices. (parsimony.lasu.dev)
AI has reached a level where good enough outputs can replace human effort across many fields, leading to a world where most people rely on basic tools rather than developing advanced skills. This shift risks diminishing the pursuit of mastery, as the effort required to become an expert becomes less necessary with accessible AI assistance. (tracingwoodgrains.com)
The video questions whether Claude's default settings are sufficient for AI-driven 3D printing. It explores the capabilities and limitations of current AI models in this application. (youtube.com)
Linus Torvalds has used an AI language model, Claude Code, to assist in writing and reviewing Linux kernel code for personal projects. He clarified that he is not using AI to develop the Linux kernel itself, but for experimental and educational purposes. (zdnet.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.