FinixDocBench - FinixDigital: leaderboard
Part of FinixDocBench (Wang et al., 2026), a financial document parsing benchmark: 500 digitally native financial report and insurance clause pages. The composite score from 0 to 100 averages text accuracy (one minus normalized edit distance), table structure similarity (TEDS) and reading-order accuracy; higher is better.
Metric: S_FDB (0-100) = (100 (1 - E_text) + TEDS_table + 100 (1 - E_order)) / 3, from the character-level normalized edit distance of the full-text Markdown, the tree-edit-distance similarity of the table structures and the normalized edit distance of the reading order; 500 digitally native financial report and insurance clause pages. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B Instruct | 87.26 | #522 |
| 2 | Kimi K2.5 | 85.05 | #256 |
| 3 | Qwen 3.5 397B A17B | 84.9 | #246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finixdocbench-finixdigital · How It Works · Data refreshed daily, snapshot 2026-11-13.