OCRia-1 is a 4-billion-parameter vision-language model built for high-accuracy, low-cost, low-latency document parsing and is deployed in LlamaParse. Across multiple independent benchmarks it sets new state-of-the-art results for models under 35B parameters: ParseBench overall 87.4 (a 12.1-point lead), ParseBench chart extraction 85.1 (19.0-point lead), and formatting preservation 83.4 (15.5-point lead). It also tops table-focused evaluations - ParseBench tables 91.5, OLMOCR-BENCH 92.9, and PULSEBENCH-TAB 84.1 - consistently outperforming competing 4B and larger open models on multilingual, long, and complex-table categories.
The model’s strengths come from architecture and a training pipeline targeted at failure modes: recovering nested headers, merged cells, irregular grids, and preserving relationships so numerical values remain correctly associated with labels; converting charts into structured HTML tables that map series to values; and retaining inline styling and math by exporting formatting and LaTeX. Training used over 100k document pages with synthetic generation, VLM-assisted bootstrapped annotations, and human review to cover dense tables, multi-page charts, mathematical notation, multilingual text, and difficult scans. That combination positions OCRia-1 for enterprise workflows - financial analysis, insurance, compliance, and automated document agents - where accurate structured extraction from complex documents is critical.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
ThinkingCap Gemma-4-26B-A4B reduces the model's thinking tokens by up to 69%, while maintaining the same accuracy across twelve benchmarks. This makes it the most efficient ThinkingCap model so far, with significant reductions in inference costs and faster responses. (bottlecapai.com)
A survey shows that only 13% of companies have achieved significant or transformational value from AI. Many organizations struggle to realize substantial returns from AI investments. (gartner.com)
An AI analyzed 45,047 Enron emails to create a detailed, citation-backed wiki about the company's knowledge and contacts. The project tracks when facts became outdated, providing a dynamic historical record. (thisandthat.chat)
Meta's Muse AI assistant has been used by over 10 million people since its launch, with high demand prompting Stripe to raise credit limits. Meta plans to monetize Muse by charging businesses a fee for sales completed through the AI service. (fortune.com)
Jev is a basic decision model that gained hype for its speed and low data requirements but lacks reasoning and multimodal features. Industry hype around Jev exposes the current stagnation and superficial innovation in AI after four years of incremental progress. (blog.alexewerlof.com)
China has invited the European Commission to join its artificial intelligence alliance. The offer aims to foster collaboration on AI development and standards. (rapporteur.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.