Open-Source 1.3B Document Parsing Model OXR Tops Three Benchmarks

The SoMark team has open-sourced OXR-1.0 (Optical Everything Recognition), a document parsing model with just 1.3B parameters that converts PDFs and document images into Markdown or JSON with layout coordinates, targeting RAG and Agent use cases. It ranks first across all three benchmarks—OmniDocBench (97.18, beating Gemini 3 Pro's 92.91), olmOCR-Bench (87.84), and DocParsingBench (87.76)—making it the current small-parameter leader in document parsing, with Infinity-Parser2 as its closest competitor.

Confirmed

• Benchmark results: OmniDocBench 97.18, olmOCR-Bench 87.84, DocParsingBench 87.76—first place in all three
• Speed: 557.8 tokens/second recognition on a single NVIDIA RTX PRO 6000, with end-to-end throughput of 0.99 pages/second—about 101 seconds for 100 pages, roughly 60 seconds faster than DeepSeek-OCR-2 on the same task
• Capabilities: supports recognition of 13 document element types including text, tables, formulas, and chemical structures; internal recognition and reading order are processed in parallel, while layout analysis runs as a separate external step that can be swapped for other layout models as needed
• Training method: billed as the first document parsing model to use recursive self-improvement (RSI) in post-training—running multiple inferences per page and flagging erroneous or inconsistent samples; having a general VLM explain the errors; clustering explanations into recurring problem types; generating new image-annotation pairs in the style of real samples; retraining and looping
• Pretrained from scratch, outperforming larger closed- and open-source rivals on multiple metrics with only 1.3B parameters

Why it matters

• Beating Gemini 3 Pro by a wide margin on OmniDocBench shows specialized small models can still compete with top-tier general-purpose LLMs on vertical document parsing tasks
• High throughput (1 page/second) plus open-source availability offers direct engineering value for RAG and Agent pipelines that need bulk document processing
• The RSI post-training loop provides a path toward automatically constructing error-correction data, which other domain-specific models could adopt

2026-10-23 ~ 2026-10-23 · 6 related posts

• [source] 1.3B Open-Source Model OXR Beats Gemini 3 Pro on Document Parsing, Tops 3 Benchmarks — JaynitMakwana · 2026-10-23
• [source] OXR Speed Details: 557.8 tok/s on One GPU, 100 Pages in 101s vs 161s for DeepSeek-OCR-2 — JaynitMakwana · 2026-10-23
• OXR Sweep Detailed: #1 on All 3 Benchmarks, 101 Seconds for 100 Pages at 1.3B Params — JaynitMakwana · 2026-10-23
• OXR tops all 3 doc parsing benchmarks, 100 pages in 101 seconds at 1.3B params — JaynitMakwana · 2026-10-23
• How OXR's RSI loop auto-labels its own failures to retrain without humans — JaynitMakwana · 2026-10-23
• [source] Open-source doc parsing model OXR hits SOTA with 1.3B params and RSI post-training — JaynitMakwana · 2026-10-23