olmOCR 2
olmOCR 2 is Ai2's open-source 7B vision model that turns PDFs and images into clean text in reading order, trained with reinforcement learning on unit-test rewards.
A release of olmOCR.
- Released
- October 2025
- Model ID
- allenai/olmOCR-2-7B-1025, allenai/olmOCR-2-7B-1025-FP8
- Size
- 7B, fine-tuned from Qwen2.5-VL-7B-Instruct, in BF16 and FP8
- License
- Apache-2.0
- Also searched as
- olmocr-2, olmocr2, olmOCR-2-7B-1025
What is new in olmOCR 2
- Fine-tuned on the olmOCR-mix-1025 dataset, then trained further with GRPO reinforcement learning to handle math equations, tables and other hard cases.
- About 4 points higher on olmOCR-Bench than the previous pipeline (82.4 against 78.5), and ahead of Marker (76.1) and MinerU (75.8) in Ai2's comparison.
- Biggest gains on hard pages: tables 84.9% (from 72.9%), multi-column layouts 83.7% (from 77.3%), old math scans 82.3% (from 79.9%).
- Handwritten and typewritten documents were added to the training data.
- Ships with olmocr toolkit v0.4.0. The 0725 and 0825 releases before it made FP8 the default and fixed auto-rotation and hallucinations on blank pages.
olmOCR 2 benchmarks
| Benchmark | Score | Note |
|---|---|---|
| olmOCR-Bench, overall | 82.4 ± 1.1 | FP8 model; BF16 scores 82.3 |
| olmOCR-Bench, tables | 84.9 | FP8 model |
| olmOCR-Bench, headers and footers | 96.1 | FP8 model |
Scores as published by the authors. We have not rerun them.
Use olmOCR 2 through an API
OCRRouter runs a different release of olmOCR as allenai/olmocr.
olmOCR versions
- olmOCR: all versions, price, speed and features on OCRRouter
- olmOCR 2 (October 2025), this page
Questions
Sources
Last updated .