1.3B open-source model OXR tops olmOCR-Bench, OmniDocBench and DocParsingBench with SOTA speed
aftahi_ai · x · 2026-10-21
SoMark released OXR, a 1.3B-parameter open-source document parsing model that claims #1 on olmOCR-Bench, OmniDocBench and DocParsingBench, scoring 97.18 vs PaddleOCR-VL-1.6's 96.34 while parsing 100 pages in 101s vs 144.7s on an RTX PRO 6000, with the largest gap on tables.
Its post-training uses a recursive self-improvement (RSI) loop, a first in document parsing, with layout modeled separately and ordering trained alongside recognition. Pretrained from scratch, OXR converts PDFs and document images into agent-friendly Markdown/JSON with layout coordinates. Code is on GitHub (MIT-style license) with a free API and online demo.
Related event: Open-Source 1.3B Model OXR Tops Document Parsing Benchmarks(4 posts)→
More from coding & agent
- A simple context hygiene trick for Claude Code: audit CLAUDE.md with claude-md-management — daniel_mac8 · 2026-10-21
- Laude Institute backs WaggleHaus, an open-source query engine for agentic queries over data lakes — lateinteraction · 2026-10-21
- komorama's weekly: the AI Peter Principle—first 90% never faster, second 90% never slower — lennysan · 2026-10-21
- Orchestrating open-source video models with Opus 5.5 cuts costs ~80x vs Seedance — mickeyxfriedman · 2026-10-21
- Multi-agent swarm architectures are being discovered, not designed — shared state emerges as the norm — JensHonack · 2026-10-21
- UC Berkeley open-sources SkyDiscover, an AI framework that discovers algorithms — CShorten30 · 2026-10-21