olmOCR 2
Allen Institute for AI
The Allen Institute for AI's document OCR model, fine-tuned from Qwen2.5-VL-7B-Instruct on the olmOCR-mix-1025 dataset of 270,000 PDF pages and then trained with GRPO reinforcement learning against deterministic unit-test rewards. It reads a page image in a single pass and emits structured text directly: Markdown for headings and document structure, HTML for tables, and LaTeX for equations. Weights ship in BF16 and FP8 alongside the training data, toolkit and fine-tuning code.
Strengths
- Reads a complex page in a single pass, with the vision encoder and decoder producing structure directly rather than stitching it together with post-hoc rules
- Emits Markdown for document structure, HTML for tables and LaTeX for equations, so the output is usable without a post-processing stage
- Scores 82.4 on olmOCR-Bench, ahead of the specialised tools Marker (76.1) and MinerU (75.8)
- Large gains on traditionally hard cases: 84.9% on tables (from 72.9%), 83.7% on multi-column (from 77.3%), 82.3% on old math scans (from 79.9%)
- Trained with GRPO reinforcement learning against pass/fail unit-test rewards, so it optimises verifiable correctness rather than approximate text similarity
- Fully reproducible release: BF16 and FP8 weights plus the olmOCR-mix-1025 and olmOCR-synthmix-1025 datasets, training code and fine-tuning scripts
- FP8 build reaches 3,400 output tokens per second on a single H100
Caveats
- Old scans is by far its weakest benchmark row at 48.3, against an 82.3 overall, so degraded historical page images remain unreliable
- Ai2 claims state of the art only for 'English-language digitized print documents', and the model card lists English as the only language
- Despite the Apache 2.0 grant the card adds that the model 'is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines'
- Input must be a single page image rendered with its longest dimension at 1288 pixels, and the prompt must carry document metadata, so in practice the olmOCR toolkit is required
- The GRPO rewards come from the same unit-test framework that olmOCR-Bench uses to score the model, though Ai2 states train and test splits are kept clean and it never trains on bench documents
- Ai2 recommends the FP8 sibling 'for all practical purposes except further fine tuning', so this BF16 repository is not the deployment target
- No context window, output limit or training cutoff is stated on the model card
- Vendor
- Allen Institute for AI
- License
- Apache-2.0
- Release Date
- 2025-10-21
- Modalities
- imagetext
| Benchmark | Score | Max |
|---|---|---|
| olmOCR-Bench (overall) | 82.3 | 100.0 |
| olmOCR-Bench - Tables | 84.3 | 100.0 |
| olmOCR-Bench - Old Scans | 48.3 | 100.0 |
Capabilities
Resources
Pricing
Open weights under Apache 2.0, free to self-host. Ai2 measures the FP8 build at 3,400 output tokens per second on a single H100, fast enough to process 10,000 pages for less than $2. Also offered through third-party APIs on DeepInfra, Parasail and Cirrascale.