OmniDocBench v1.5 (Meta two-component scoring)
Multimodal · 2026-10-17
Meta's internal scoring of the 1,355 OmniDocBench v1.5 pages in the Muse Glimmer report. It folds formulas into the edit-distance text groups and reports a two-component weighted average of text and tables instead of the official text, table TEDS and formula CDM composite, and it matches elements with the Hungarian algorithm rather than the official adaptive matching. It is kept separate from the official Overall score.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.6 27B | 77.8 |
| Muse Glimmer 30B | 75.8 |
| Gemma 4 31B IT | 72.5 |