How LlamaIndex launched ParseBench-2 on X
Benchmark that scores document parsers by error type and metric alignment
What LlamaIndex launched

Today we’re introducing ParseBench-2 - our most ambitious document parsing benchmark to date. It is the only benchmark that provides both extreme granularity into every single type of production parsing error, while also making sure that metric scores are highly aligned with https://t.co/TUFFGze8Kc
Transcript
How LaunchVideo.com rated the launch
LaunchVideo.com rating
About the launch
LlamaIndex builds document parsing and workflow tools so agents can work from real PDFs, scans, tables, and forms. Cofounder and CEO Jerry Liu released ParseBench-2 on November 12, 2026, an open benchmark for teams choosing a parser for production. The first ParseBench, published in April, had already become a reference point, with the company reporting more than 50,000 dataset downloads a month and citations from teams at NVIDIA, Cohere, and Perplexity. This update targets a specific failure in that earlier scoring. On a 459-cell table, swapping a number cost 0.13 points out of 100, while a correct table returned as Markdown instead of HTML scored zero. ParseBench-2 grows the set from about 2,100 cases to 4,317 page-level test cases drawn from 2,645 public documents, and adds forms beside text, formatting, tables, charts, and visual grounding. The team wrote more than 330 controlled error scenarios, rated each critical, major, minor, or harmless, and retuned deterministic metrics so the penalty tracks that severity. The swapped number in the example now costs 10 points, and the Markdown rendering costs nothing. Chart scoring covers all 36,958 values across 1,025 charts, and form checks require values, checkbox states, and signatures to stay with the right field, including blanks that must stay blank. The dataset and evaluation code are public under Apache 2.0, a research paper is available, and outside parsers can submit results. LlamaIndex ranked 41 configurations on the new leaderboard. LlamaParse Agentic Plus led at 89.8 overall, ahead of Claude Opus 5.5 at 85.7 and Reducto in agentic mode at 83.9. The same release introduces OCRia-1.5, the model family now running LlamaParse Cost Effective, Agentic, and Agentic Plus, with the company’s stated gains concentrated on tables, forms, and layout. Scores from the first ParseBench are not comparable with these results.
Tags
Questions
Frequently asked questions
The LlamaIndex launch post reached 109K views on X, with 214 likes, 28 reposts, 108 comments, 58 bookmarks.
LlamaIndex launched on X on November 12, 2026 with the message "Benchmark that scores document parsers by error type and metric alignment".
LlamaIndex builds document parsing and workflow tools so agents can work from real PDFs, scans, tables, and forms.
LaunchVideo.com rated the LlamaIndex launch 2.80 out of 5: hook 1/5, editing / creativity 3/5, copy 4/5, sentiment of launch 4/5, distribution strategy 2/5.







