Building It
WHERE FOUNDERS SHIP /// WHERE FOUNDERS SHIP /// WHERE FOUNDERS SHIP /// WHERE FOUNDERS SHIP /// WHERE FOUNDERS SHIP /// WHERE FOUNDERS SHIP ///
AnalysisBuild time: 3 min

SoMark Releases OXR, a 1.3B Open-Weight Document Parsing Model That Ranks First on Three Public Benchmarks

SoMark released the code and weights for OXR, a 1.3B document parsing model that ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench in SoMark's comparison.

Shipped by Adrian WeiOctober 21, 2026
SoMark Releases OXR, a 1.3B Open-Weight Document Parsing Model That Ranks First on Three Public Benchmarks

SoMark has released the code and weights for OXR (Optical Everything Recognition), a 1.3-billion-parameter model that converts PDFs and document images into Agent-Friendly data in Markdown or JSON, with layout coordinates attached to every element.

The release, which went live on October 20 at 9 a.m. Pacific, is aimed at teams building retrieval-augmented generation (RAG) systems, AI agents and document workflows, where a parsing error made at ingestion travels through every step that follows.

OXR, from the SoMark team.

OXR, from the SoMark team.

Accuracy across three benchmarks

According to SoMark's published comparison, OXR scores 87.84 on olmOCR-Bench, 97.18 on OmniDocBench and 87.76 on DocParsingBench, placing first on all three. The company notes that peer-model scores come from published papers, model cards and the official leaderboards rather than from its own re-runs, and that the evaluation pipeline is public so others can reproduce the results. On DocParsingBench, OXR leads in the overall, text, formula and table categories among 17 compared systems.

The model recognizes 13 document element types, including tables, formulas, chemical equations, chemical structures and code blocks. Tables come out as HTML or Markdown, formulas as LaTeX, chemical equations in mhchem notation and chemical structures as SMILES. SoMark says the model also handles hard layouts such as A3 exam papers, newspapers, vertical text and handwriting.

Overall scores on olmOCR-Bench, OmniDocBench and DocParsingBench. Source: SoMark.

Overall scores on olmOCR-Bench, OmniDocBench and DocParsingBench. Source: SoMark.

Speed on a single GPU

SoMark also measured end-to-end parsing speed for 21 models on the same frozen 100-page subset of OmniDocBench, each on one NVIDIA RTX PRO 6000 Blackwell GPU. OXR averaged 100.972 seconds per 100 pages, about 0.99 pages per second. The next fastest, OvisOCR2, took 134.778 seconds, which puts OXR's throughput 33.5% higher.

SoMark attributes both the accuracy and the speed to the model architecture and its RSI post-training loop working together. On the architecture side, layout analysis runs as a separate model, while recognition and reading-order prediction run together inside OXR, which removes a sequential reading-order stage. The training side is described below.

Pluggable layout analysis and RSI

Because layout analysis is separate, teams can swap in their own layout model when their documents need different element categories, region boundaries or grouping rules, and keep OXR for recognition and reading order.

For post-training, SoMark uses what it calls a recursive self-improvement (RSI) loop. A general vision-language model reviews OXR's outputs, disagreement across repeated predictions flags uncertain samples, and a clustering model groups the likely causes. The loop then plans the next training round and generates targeted training data, with rule-based checks in place of manual labeling. To the team's knowledge, OXR is the first document parser trained this way.

Availability and license

The weights are on Hugging Face and ModelScope, and the code is on GitHub. Running locally requires Linux and an NVIDIA GPU with BF16 support; 24 GB of VRAM is enough for low-concurrency testing, and 48 GB or more is recommended for long documents. A free hosted API and online demo require no signup or API key, though SoMark says requests are queued and the service is meant for testing, not production.

OXR ships under a custom license based on Apache 2.0. Commercial use is free for organizations below USD 2 million in total funding or annual revenue; above either threshold a separate commercial license is required, and the license also restricts offering a competing hosted API. A technical report is listed as coming soon.