🚀PP-OCRv6 is officially released！ 🔥PaddleOCR’s new OCR model series scales from 1.5M to 34.5M parameters, bringing stronger accuracy, faster inference, and broader deployment options — from browsers and edge devices to servers. 📊What’s new: 🔸Tiny / Small / Medium models: 1.5M, 7.7M, 34.5M params 🔸+4.9% detection accuracy and +5.1% recognition accuracy over PP-OCRv5 🔸Up to 5.2× faster CPU inference with OpenVINO 🔸50 languages in one unified model 🔸New scenarios: PCB, CAD drawings, digital tubes, dot-matrix text 🔸Apache 2.0 open source ✨Lightweight OCR, built for the AI data era. 🔗Try it: 🌐 paddleocr.com 💻 github.com/PaddlePaddle/Padd… 🤗huggingface.co/collections/P… #PaddlePaddle #PaddleOCR #OCR #AI #ComputerVision #OpenSource #EdgeAI

FlashMask: Faster, More Flexible Attention for Long-Context Training We’re excited to share FlashMask, a high-performance Attention library for LLM training! 🔸 Flexible support for complex attention masks 🔸 Up to 1.55× faster than FlashAttention-4 🔸 Efficient context parallelism with load balancing and computation–communication overlap 🔸 PaddlePaddle & PyTorch support via CuTe DSL 🔸 SM90 & SM100 support 🔸 Large head dimensions: 192/128, 256/256, 512/512, 576/512 FlashMask-3 is accepted at NeurIPS 2026! 🌐 GitHub: github.com/PaddlePaddle/Flas… 💻 Paper: arxiv.org/pdf/2610.17901 🤗 Project: enigmatisms.github.io/flashm…

📣 The PaddleOCR Application Case Call is now open! Whether you represent an enterprise, developer team, university or research institution, or ecosystem partner, if you have used PaddleOCR to address a real-world challenge, we invite you to share your project and experience. Your story deserves to be seen by a wider audience. Selected cases will receive higher usage quotas for PaddleOCR’s online API, promotion through official channels, opportunities to participate in technical events, and access to PaddleOCR technical exchange and ecosystem collaboration. 🗓️ Submission period: September 4–15 📩 How to participate: Click the link paddle.wjx.cn/vm/wFk5UCx.asp… to complete the questionnaire and join the group. The case template will be provided in the group. Please send your completed case materials to [email protected].

🌍Real impact, real change. 👏From smart mobility to no-code creation to cultural heritage preservation—these cases show how AI can create meaningful value across a wide range of fields.

👏Excited to see Unlimited-OCR now running in vLLM! Huge thanks to the @vllm_project community for the support and collaboration in bringing efficient long-context OCR to more developers. Check out the recipe and give it a try 👇

PaddleFormers 1.2 is officially out — DeepSeek-V4 training now supported! 🔥 128K+ long-context training, highly efficient and fully open-source. 📊 Highlights: 🔸 DeepSeek-V4 training support: the industry's most complete open-source training solutions. 🔸 Large-scale training features: Context Parallel, Packing, Document Mask Attention & Muon optimizer. 🔸 Performance boost: ultra-fused mHC / CSA / HCA operators, DeepEP / HybridEP communication, lossless FP8 training & AutoSubbatch memory balancing. ✨ Built for large-scale LLM training — longer context, higher efficiency, better optimization. 🔗 Check it out: github.com/PaddlePaddle/Padd… #PaddlePaddle #PaddleFormers #PaddleFleet #DeepSeekV4 #LLMTraining #LongContext

🧵PP-OCRv6 Tech Deep Dive Ep.4：3.9x Faster on CPU, 0.13s per Image on A100 — PP-OCRv6 Deployment & Model Selection Guide How fast can OCR really get outside the lab? PP-OCRv6 Tech Deep Dive Ep.4 answers with full end-to-end benchmarks across A100, V100, Intel Xeon CPU, and Apple M4, using PaddlePaddle, ONNX Runtime, OpenVINO, and TensorRT. The highlights: 🔸 0.13s/image on A100 with PP-OCRv6_tiny. 🔸 5.2× faster on Intel CPU: PP-OCRv6_medium vs PP-OCRv5_server with OpenVINO. 🔸 3.9× faster on Intel CPU: PP-OCRv6_tiny vs PP-OCRv5_mobile with OpenVINO. 🔸 0.35s/image on Apple M4 with PP-OCRv6_tiny + ONNX Runtime. 🔸 50 languages in one unified Medium/Small model. 🔸 88.4% English accuracy and 88.0% Latin-script accuracy with PP-OCRv6_medium. Deployment guide: 🔹 High-concurrency API? Choose Medium. 🔹 CPU document systems? Choose Small. 🔹 Mobile or embedded devices? Choose Tiny. 🔹 Multilingual business? Choose Medium or Small. Across the full series, PP-OCRv6 shows one thing clearly: in dedicated OCR tasks, lightweight architecture + high-quality training data can be more practical than simply scaling parameter count. Architecture, detection, recognition, deployment — the PP-OCRv6 technical deep dive is now complete. #PaddleOCR #PPOCRv6 #OCR #Deployment #OpenVINO

🧵PP-OCRv6 Tech Deep Dive Ep.3: How CTC + NRTR Dual-Head Architecture Prevents Hallucination in Text Recognition — PP-OCRv6 Recognition Module Deep Dive When VLMs read text, they may guess, OCR can't afford that. PP-OCRv6 Tech Deep Dive Ep.3: after detection locks the location, recognition must read it right. In hallucination tests, best VLM reaches only 85.0%, while PP-OCRv6_medium hits 93.2%. The recognition design is built on visual evidence, not language priors: 🔸 LightSVTR neck: 1×7 depthwise conv scans local text structures before global attention — +0.47% accuracy. Additive skip connections replace Concat, reducing parameters and stabilizing training. 🔸 CTC + NRTR dual-head training: CTC decodes purely from visual features at inference — no language prior, no hallucination. NRTR joins only during training, providing language regularization at zero inference cost — +1.16% accuracy. 🔸 Knowledge distillation boosts Tiny from 71.83% to 74.52% — recovering the capacity loss from removing the neck. Result: 83.2% recognition accuracy across 15 scenarios, surpassing PP-OCRv5_server (78.1%), with especially strong gains on Japanese (+16.8%), ancient Chinese (+12.0%), and screen text (+14.4%). 💬 What's your deployment environment — cloud GPU, on-prem CPU, or edge device? 📅 Next in the final episode (Ep.4): 3.9x Faster on CPU, 0.13s per Image on A100 — PP-OCRv6 Deployment & Model Selection Guide #PPOCRv6 #OCR #TextRecognition #AntiHallucination #VLM

👏Great to see PP-OCRv6 land on @huggingface 👍Better accuracy, plus transformers & ONNX Runtime backends. Definitely worth a try！ #OCR #OpenSource #HuggingFace

🧵PP-OCRv6 Tech Deep Dive Ep.2: Text Detection Demystified: Precise Localization for Small, Curved, and Industrial Text — PP-OCRv6 Text Detection Text detection is the key to the success or failure of OCR. In benchmark tests, general-purpose VLMs still underperformed: Gemini-3.1-Pro's detection Hmean was only 46.8%, while GPT-5.5 dropped to 38.3%. In contrast, PP-OCRv6_medium achieved a detection Hmean of 86.2%. PP-OCRv6 Tech Deep Dive Ep.2 breaks down the "three-pronged" strategy behind its detection module: 🔸 RepLKFPN replaces the 3×3 kernels in RSEFPN with 7×7 large kernels, expanding the receptive field while reducing FPN neck parameters from 172K to 118K — a 31% reduction. 🔸 Auxiliary deep supervision adds training-only heads on P2, P3, and P4, providing direct gradient signals to intermediate layers to learn discriminative edge and texture features. 🔸 Focal Loss works in tandem with Dice Loss — Dice aligns overall shape, while Focal Loss mines hard pixels — contributing +1.15% Hmean in ablation. From curved text on tires to dense small text on electronic screens, dot-matrix text, and cards, PP-OCRv6_medium delivers more comprehensive and accurate localization. 💬 Does your recognizer ever misread numbers or hallucinate on low-res text? 📅 Next in Ep.3: How CTC + NRTR Dual-Head Architecture Prevents Hallucination in Text Recognition — PP-OCRv6 Recognition Module Deep Dive #PPOCRv6 #TechDeepDive #OCR #TextDetection #VLM

🧵PP-OCRv6 Tech Deep Dive Ep.1: In the Era of Large Models, Why Does Lightweight OCR Still Have Irreplaceable Value? — PP-OCRv6 Architecture Design Can a 34.5M-param OCR model challenge 100B-scale VLMs? PP-OCRv6 Tech Deep Dive Ep.1 explains why lightweight OCR still matters in the large-model era. In real-world OCR scenarios, VLMs still struggle with inaccurate localization, hallucinations, and high inference costs. So PP-OCRv6 rebuilds the backbone around LCNetV4: 🔸 MetaFormer-style design: Token Mixer answers “where is the text?”, while Channel Mixer answers “what is the text?” 🔸 Structural re-parameterization: multi-branch training, fused 3×3 DWConv inference, with no extra overhead or accuracy loss 🔸 One backbone, two task modes: a 2D feature pyramid for detection, and an asymmetric stride (2,1) for recognition 🔸 Three model specs: Tiny for edge CPU devices, Small for balanced deployment, Medium for industrial high-accuracy pipelines With this design, PP-OCRv6_medium reaches 86.2% detection Hmean and 83.2% recognition accuracy, fully surpassing PP-OCRv5_server while running faster. 💬 Where does your OCR pipeline struggle most—tiny text, curved text, or edge speed? 📅 Next in Ep.2: Text Detection Demystified: Precise Localization for Small, Curved, and Industrial Text — PP-OCRv6 Text Detection #PPOCRv6 #OCR #MetaFormer #PaddleOCR #VLM

Thanks for the shoutout! As @slimcat0101 mentioned, we believe PP-OCRv6 marks a new milestone for both us and the OCR community. A huge thank you to the community🥰. We're always open to your feedback.🙌

🚀 PaddleOCR 3.7 is here — ONNX Runtime now supported! 🔥 One parameter to switch inference backends. Zero code changes. 📊 What's new: 🔸ONNX Runtime backend: cross-platform, lightweight, multi-hardware acceleration. 🔸Multi-hardware support: CPU via OpenVINO, GPU via CUDA & TensorRT. 🔸PP-OCRv6 also onboard: Tiny / Small / Medium — up to 3.9× faster on CPU. ✨ From servers to edge devices — flexible deployment, your choice of engine. 🔗 Try it: 🌐 paddleocr.com 💻 github.com/PaddlePaddle/Padd… 📖 paddleocr.ai/latest/versio #PaddlePaddle #PaddleOCR #OCR #ONNX #ONNXRuntime

👏Surprise: free to use — give it a try！ Baidu CoBuddy is now live on @novita_labs — A powerful code-focused model for developers and AI agents 🚀 ✨131K context · 65K output · native tool calling

🔥 Official Launch: PaddleOCR & ERNIE Image Are Now Available as Dify Plugins — Powering the Next Wave of AI Agents We’re excited to see PaddleOCR and ERNIE Image now available as official Dify plugins — bringing document intelligence and image generation directly into Dify’s agentic AI workflows. A big thank you to @dify_ai , our valued PaddleOCR OCEAN Alliance member, for the continued ecosystem collaboration. 🤝 1️⃣ PaddleOCR — Document AI for agent workflows Turn images, scanned PDFs, and multilingual documents into clean structured data. 🔹 Powered by PP-OCRv5, PP-StructureV3 & PaddleOCR-VL 🔹 High-precision OCR and document parsing 🔹 Structured outputs for downstream chunking, vectorization, and RAG 🔹 Supports private/on-prem deployment for enterprise scenarios 👉 marketplace.dify.ai/plugins/… 2️⃣ ERNIE-Image — Image generation inside Dify Generate visual assets directly in your Dify workflows. 🔹 Free image generation 🔹 Turbo mode with 8-step inference 🔹 OpenAI-style API for easy integration 🔹 Built for posters, social media visuals, and creative production 👉 marketplace.dify.ai/plugin/l… No extra deployment. Add the plugin node, connect your workflow, and start building. ✨ 💻 Learn more: PaddleOCR 👉 github.com/PaddlePaddle/Padd… ERNIE Image 👉 aistudio.baidu.com/ernieimag… Dify 👉 github.com/langgenius/dify #PaddleOCR #ERNIEImage #Dify #DocumentAI #AIGC #AgenticAI #OpenSource

🚀PaddleOCR-VL 1.6 Officially Released! We are thrilled to announce the official release of PaddleOCR-VL 1.6 — this version has set a new SOTA record of 96.33% on OmniDocBench, outperforming both open-source and proprietary solutions in text, formula, and table recognition. 📊 Key Highlights: 🔸 Ranked #1 on OmniDocBench v1.5 and Real5-OmniDocBench as well 🔸 Significant improvements in table, classic text, and rare character recognition 🔸 Enhanced seal, spotting, and chart recognition 🔸 Fully compatible with v1.5 architecture — zero migration, plug-and-play From financial contracts and legal documents to research reports and historical archives — empower your document intelligence workflows by providing high-quality data to large language models (LLMs) and retrieval-augmented generation (RAG) systems. ✨ Industry-leading accuracy. Zero migration. Plug-and-play. 🔗 Read more: github.com/PaddlePaddle/Padd… #PaddlePaddle #PaddleOCR #AI #ComputerVision

Congratulations to NabuOCR 🎉 Thanks Zack for trusting us and building something truly meaningful with PaddleOCR. From ancient cuneiform tablets to AI-powered text recognition — technology's journey across millennia is what makes it so inspiring.