Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresParseBench-2 (LlamaIndex) scores document parsing on 4,317 page-level test cases from 2,645 enterprise documents in six equally weighted dimensions (text, formatting, tables, charts, visual grounding and forms), from 0 to 100. The board compares hosted parsing products and dedicated OCR models, which Atlas doesn't track, with direct prompts to vision-language models; only rows for models Atlas carries are imported. LlamaIndex's own LlamaParse leads the board, so it is shown as a diagnostic outside the capability fit.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | not reported | 85.8 | |
| 2 | Gemini 4 Argon | low | 84.2 | |
| 3 | GPT-5.6 Terra | low | 77.1 | |
| 4 | Gemini 3 Flash Preview | minimal | 76.6 | |
| 5 | Gemini 3.8 Flash | low | 75.5 | |
| 6 | Gemini 3.6 Flash | — | 74.7 | |
| 7 | Muse Ember | low | 74.5 | |
| 8 | GPT-6.1 Terra | low | 74.3 | |
| 9 | Claude Sonnet 5.5 | not reported | 73.8 | |
| 10 | GPT-6 Luna | low | 73.5 | |
| 11 | Claude Haiku 5.5 | not reported | 70.1 | |
| 12 | GPT-5.6 Luna | low | 69.1 | |
| 13 | Gemini 3.5 Flash-Lite | — | 61.4 | |
| 14 | Qwen3.5-4B | disabled | 42.3 | |
| 15 | Gemma 4 E4B IT | not reported | 42.3 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).