SoMark open-sources OXR, a 1.3B document parsing model trained with a recursive self-improvement loop
rohanpaul_ai · x · 2026-10-28
SoMark 团队开源了文档解析模型 OXR（GitHub 1.7k star），把 PDF 和文档图片转成带布局坐标、对 Agent 友好的 Markdown/JSON 数据，从零预训练，宣称精度与速度均达 SOTA。
- 架构思路：小布局模型负责给每个区块画框，1.3B 的 OXR 读取所有框并设定阅读顺序，布局模块可按领域替换。
- RSI 后训练循环：由通用 VLM 标记 OXR 出错或不确定的页面，聚类归组失败原因，再针对每组自动生成新的带标注训练页——全程无需人工标注，实现递归自我改进。
- 提供 Docker 部署、免费 API 和在线 demo，附中英文文档。
Related event: SoMark Open-Sources OXR-1.0, a 1.3B Document Parsing Model Hitting SOTA(2 posts)→
More from Models
- Tencent's Hy-SimulMT-30B-A3B-FP8 translation model trends on Hugging Face — tencent · 2026-10-28
- Google reportedly launches Genie 4, its strongest world model yet — oran_ge · 2026-10-28
- Grok Bot reportedly building Discord support, Slack and Teams bots coming — nima_owji · 2026-10-28
- Kimi K3.1 scores 38% on RareBench, matching GPT-6 Astra at a quarter of the cost — danielmckinn0n · 2026-10-28
- 48 local models tested as agents on M1 Max 64GB: qwen3.8-27b tops real-world score — HAVT_ · 2026-10-28
- GPT-6 Astra hits SOTA on ARC-AGI-3, inventing its own programming language — typewriters · 2026-10-28