OXR 1.0

Name: OXR 1.0
Author: SoMark

SoMark · released 2026-10-02 · data checked 2026-11-06

OXR 1.0 is a 1.5B-parameter OCR and document-parsing model from SoMark (CN), released 2026-10-02 under the OXR Open Source License license. It scores 97.03 on OmniDocBench v1.6 (#1 of 33) and 87.8 on olmOCR-Bench (#1 of 24). It needs about 4.1 GB of memory at 16-bit.

Trained from scratch. Outputs Markdown or JSON with layout boxes and 13 element types, including chemical structures.

OmniDocBench v1.6

97.03

#1 of 33 models

olmOCR-Bench

87.8

#1 of 24 models

Parameters

1.5B

OCR specialist

Memory to run

~4.1 GB

16-bit weights plus about 1 GB

Benchmarks

OmniDocBench v1.6 Leaderboard97.03

Best score among the models listed.

olmOCR-Bench Lab-reported87.8

Best score among the models listed.

The tick marks the best score on each benchmark. See how to read the ranking.

Specifications

Lab: SoMark (CN)
Type: OCR / document parsing specialist
Total parameters: 1.5B
License: OXR Open Source License
Custom license based on Apache 2.0 with additional terms; reported to require a paid license above $2M revenue. Check the LICENSE file before commercial use.
Memory to run: ~4.1 GB at 16-bit
Released: 2026-10-02

Weights and usage on Hugging Face ↗

Closest alternatives

Youtu-Parsing-Omni (5.3B, 96.96 on OmniDocBench, Youtu-Parsing license)
TeleOCR (1.4B, 96.91 on OmniDocBench, Apache-2.0)
Xiaomi-OCR-0 (870M, 96.83 on OmniDocBench, Apache-2.0)
OvisOCR3 (850M, 96.74 on OmniDocBench, Apache-2.0)

Sources

Something wrong here? Report it.