Is OXR-1.0 state of the art?Pending benchmarks.Announced Oct 21, 2026; no independent results yet.
SoMark's open-weight document parsing model. It turns PDFs and document images into Markdown or JSON with layout coordinates, recognising 13 element types including tables, formulas, chemical equations and structures, stamps and code blocks. Layout analysis is a separate, swappable model. Weights under SoMark's OXR License, which adds commercial thresholds to Apache 2.0.
The details so far
- Lab
- SoMark
- Announced
- Oct 21, 2026
- API
- Live
- Open weights
- Published
- Parameters
- 1.3B (SoMark); the Hugging Face safetensors file totals about 1.53B
- Task
- Document parsing (PDF and images to Markdown / JSON), with a pluggable layout-analysis model (D-FINE based, shipped with the weights)
- Training
- Pretrained from scratch; post-training uses a recursive self-improvement loop on failed and weak samples
- License
- OXR License: Apache 2.0 plus overriding terms. Commercial use is free only below USD 2M total funding and USD 2M annual revenue; online services must credit SoMark OXR; competing hosted parsing/OCR APIs are not allowed
- Weights
- Hugging Face (SoMarkAI/OXR-1.0) and ModelScope (SoMark/OXR-1.0), BF16; repos created Oct 2–5, 2026
- Hosted access
- Free API and online demo with no key or sign-up, no stability guarantee
- Serving
- vLLM via an out-of-tree plugin; 24 GB GPU minimum, 48 GB+ recommended
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: first in SoMark's own comparisons on olmOCR-bench (87.84), OmniDocBench (97.18) and DocParsingBench (87.76), with peer scores taken from published papers, model cards and leaderboards.
Lab claim: fastest of 21 models in SoMark's speed test at 557.77 native decode tokens/s and 0.99 pages/s on one RTX PRO 6000 Blackwell GPU.
Sources
REPOOXR-1.0 on ModelScopemodelscope.cn/models/SoMark/OXR-1.0 · Oct 21, 2026REPOOXR-1.0 on Hugging Facehuggingface.co/SoMarkAI/OXR-1.0 · Oct 21, 2026REPOOXR on GitHub (code, license and benchmark notes)github.com/SoMarkAI/OXR · Oct 21, 2026PRESSModelScope announcement on Xx.com/ModelScope2022/status/2112901338178687182 · Oct 21, 2026Tracking continues until an independent board lists it.