Open-Source 1.3B OXR Tops Three Document Parsing Benchmarks

Document parsing remains a difficult layer of the AI stack. Extracting characters from a PDF is only the beginning: a useful system must also recover reading order, table relationships, formulas, captions and the links between text and images. If a document is converted into scrambled text, downstream retrieval-augmented generation and agent workflows inherit faulty context regardless of how capable the language model is.

OXR is an open-source project that expands the goal of OCR toward recognition of “everything” in a document. According to its developers, the model has about 1.3 billion parameters and handles tables, formulas, code blocks, images, chemical structures, footnotes, headers, footers and stamps. It can convert documents into structured Markdown and formulas into LaTeX. The project reports scores of 97.18 on OmniDocBench, 87.84 on olmOCR-Bench and 87.76 on DocParsingBench, ranking first on all three. These figures come from the project’s published materials and should be checked against the exact benchmark versions, hardware and inference settings.

Key points include:

• A focused small model: OXR is not positioned as a general-purpose language model. It concentrates on recognition and structural reconstruction for documents.
• RSI-based post-training: The system processes large collections of unlabeled documents, then uses a general vision-language model to classify outputs as correct, wrong or uncertain. Multiple rollouts are used to detect unstable predictions.
• Cause-based data generation: Failed samples are diagnosed and grouped by causes such as merged-cell errors, handwriting failures or sensitivity to visual perturbations. New training examples are then generated for specific weakness clusters.
• A modular pipeline: Layout analysis is separated from recognition and reading-order prediction, allowing developers to replace the layout component for specialized business needs.

The interesting part of this approach is not simply asking a model to grade itself. It connects error discovery, diagnosis, data construction and retraining into a feedback loop. Document parsing is relatively suitable for this setup because many outputs have explicit structural constraints. However, automatically generated data can also amplify evaluator mistakes, so human inspection and independent testing remain important for formulas, complex tables and long-tail layouts.

The project also emphasizes throughput. Its developers report that a 100-page document can be parsed end to end in about 101 seconds on a single RTX PRO 6000. The architecture combines a lightweight layout model with a main model that handles recognition and reading order, aiming to avoid the long autoregressive outputs that slow down fully generative document systems. Code, weights, evaluation scripts, an online demo and a free API are reportedly available, with Docker-based deployment options.

OXR’s broader significance is the shift from “turning images into text” to turning documents into structured assets that AI systems can use. If the reported results prove reproducible, this approach could lower the barrier for importing research papers, enterprise archives and operational records into AI workflows. Developers should nevertheless evaluate privacy, service reliability, Chinese-language performance and throughput on their own hardware before production adoption.

Source: QbitAI