Smaller size and higher score are better. The frontier includes points that no other comparable point matches or improves on in both dimensions, with a strict improvement in at least one. Comparisons use source precision; displayed scores are rounded to one decimal. Parameter counts are those in the tables, including nominal backbone sizes where specified.

Overall uses five categories for ParseBench and includes headers/footers for olmOCR. All points match the selected table rows; LightOnOCR-3 uses the candidate-1 evaluations for 0.8B, 1B, and 4B. Postprocessing applies only to olmOCR; ParseBench and fr-bench-pdf2md outputs are evaluated as submitted. The same candidate checkpoints are used across all three benchmarks. Missing scores or sizes are omitted. Models with undisclosed parameter counts (such as MistralOCR4.1) cannot be placed on a size axis. * LightOnOCR-2 uses its published olmOCR score excluding headers/footers; it is shown but excluded from the frontier. Mistral OCR 4 is a horizontal score reference because its size is undisclosed; it is excluded from the size frontier. Its olmOCR score is the user-provided reported 85.20; ParseBench and fr-bench-pdf2md retain their existing selected scores.

Leaderboard source revision: 3576c5240602588f596f8790a3a7161a524e59ad. Values are pinned to this snapshot.