Son Dakika
KR호날두, 월드컵 6회 연속 출전·득점 신기록 세우고도 16강서 탈락DEFruchtbarkeitsprobleme bei Frauen nehmen weltweit zuTRBM İnsan Hakları Konseyi, Sudan'daki İhlalleri Soruşturacak Misyonu OnayladıDEUSA: Millionen Menschen verlieren Krankenversicherungsschutz nach Wegfall von SubventionenARإسبانيا تتخطى البرتغال وتتأهل لربع نهائي كأس العالم 2026 بسجل تاريخيCN強烈颱風「巴威」逼近 預計週五傍晚籠罩台灣北部及東部GLOBALGianni Infantino's Fifa Presidency: A Decade of Controversy and a Potential Tipping PointGLOBALMaine Senate Candidate Graham Platner Faces Sexual Assault Allegation, Calls for Withdrawal GrowAUKaniva Gains Mains Drinking Water Supply After Century of Reliance on Rainwater and Bore WaterBRCasa da Mulher Brasileira em Manaus segue sem previsão de inauguraçãoKR호날두, 월드컵 6회 연속 출전·득점 신기록 세우고도 16강서 탈락DEFruchtbarkeitsprobleme bei Frauen nehmen weltweit zuTRBM İnsan Hakları Konseyi, Sudan'daki İhlalleri Soruşturacak Misyonu OnayladıDEUSA: Millionen Menschen verlieren Krankenversicherungsschutz nach Wegfall von SubventionenARإسبانيا تتخطى البرتغال وتتأهل لربع نهائي كأس العالم 2026 بسجل تاريخيCN強烈颱風「巴威」逼近 預計週五傍晚籠罩台灣北部及東部GLOBALGianni Infantino's Fifa Presidency: A Decade of Controversy and a Potential Tipping PointGLOBALMaine Senate Candidate Graham Platner Faces Sexual Assault Allegation, Calls for Withdrawal GrowAUKaniva Gains Mains Drinking Water Supply After Century of Reliance on Rainwater and Bore WaterBRCasa da Mulher Brasileira em Manaus segue sem previsão de inauguração
Newsgather
GeriBaidu's PaddleOCR-VL-1.6 Achieves World-Leading Performance in Document Understanding
Baidu's PaddleOCR-VL-1.6 Achieves World-Leading Performance in Document Understanding
Teknoloji
中国新闻网02.06.2026Teknoloji3 dk okumaChina

Baidu's PaddleOCR-VL-1.6 Achieves World-Leading Performance in Document Understanding

Hızlı Bakış

Baidu's new PaddleOCR-VL-1.6 model, derived from the Ernie large model, has achieved a 96.33% accuracy on OmniDocBench v1.6, surpassing competitors and ranking first globally in complex document understanding and real-world scenario parsing.

Yapay zekâ özeti

Neden Önemli?

Baidu has released PaddleOCR-VL-1.6, an advanced OCR model derived from its Ernie large model. This new version has achieved top scores in benchmark tests for complex document understanding and real-world scenario parsing.

Yazı boyutu

近日,百度正式发布 PaddleOCR-VL-1.6。作为文心衍生模型,PaddleOCR-VL-1.6在 OmniDocBench v1.6权威评测中准确率突破96.33%,在 Real5-OmniDocBench、OmniDocBench v1.5上也取得了最领先的分数,综合性能全球第一,刷新业界SOTA,在复杂文档理解和真实场景解析能力方面进一步突破。

在OmniDocBench v1.6上,PaddleOCR-VL-1.6在多项能力中实现SOTA

在权威评测集 OmniDocBench v1.6上,PaddleOCR-VL-1.6总指标达到96.33%,超越Gemini-3-Pro、GPT-5.2、MinerU-2.5-Pro、GLM-OCR等,综合性能全球第一,领跑全球通用大模型和专用OCR模型;在面向真实复杂场景构建的 Real5-OmniDocBench评测中,PaddleOCR-VL-1.6总指标达到93.19%,较 Gemini-3-Pro提升近4个百分点,在扫描件、弯折文档、屏幕拍照、光照变化及倾斜文档等五大真实场景下均保持领先表现。

测试结果显示,PaddleOCR-VL-1.6在文本、公式、表格等核心识别能力上全面领先当前主流开源及闭源方案,在表格、古籍、生僻字识别等复杂场景能力显著提升,在印章、Spotting、图表识别等多项关键能力同步增强,可更好满足文档数字化的需求。

据了解,PaddleOCR基于文心大模型训练而来,是文心大模型多模态能力的重要部分,支持超100种语言识别,用户覆盖170多个国家和地区。此次发布的 PaddleOCR-VL-1.6在 PaddleOCR-VL-1.5基础上进一步升级,通过模型驱动的数据构建机制和渐进式训练优化,在保持0.9B轻量化架构的情况下,模型准确率和复杂场景适应能力进一步提升。由于两代模型结构一致,开发者和企业用户无需进行额外适配,即可平滑迁移。

近年来,PaddleOCR持续推进文档理解能力升级,先后推出 PaddleOCR-VL、PaddleOCR-VL-1.5等多款模型。其中,PaddleOCR-VL-1.5创新支持异形框定位,在真实文档场景中展现出较强解析能力。此前,PaddleOCR GitHub Star数已突破79.2K,超过谷歌开源OCR项目 Tesseract OCR,成为全球最受开发者欢迎的开源OCR项目之一。

目前,PaddleOCR-VL-1.6已上线 PaddleOCR官网,支持网页端和API调用。同时,模型代码及权重已同步开源至 GitHub和 Hugging Face,面向全球开发者开放使用。

Açık Sorular

  • What are the specific technical innovations that led to the performance leap in PaddleOCR-VL-1.6?
  • How does PaddleOCR-VL-1.6 compare to other multimodal AI models in broader applications beyond OCR?
  • What are the future development plans for the PaddleOCR series?
  • What is the specific impact of this advancement on industries relying on document digitization?

İlgili Konular

Bu haber ilk olarak şurada yayınlandı: 中国新闻网.

İlgili Haberler

Bu konuda daha fazlaBaidu