OCR and document AI
- 9 days ago @VikParuchuri
Benchmark of sub-1B parameter OCR models on OmniParseBench
A post from @VikParuchuri says the author's team benchmarked recent small OCR model releases with under 1B parameters on OmniParseBench. The post links to no primary source or results page.
Gives engineers a reference point for choosing small OCR models, though the post itself lacks details.
- 9 days ago @vanstriendaniel
Claude Opus edits OCR model weights to stop repetition loops
The author reports asking Claude Opus to find a small OCR model and edit its weights to fix repetition loops. It reportedly changed 0.0085% of the weights with no training, in under 3 hours and about $10 of Hugging Face Jobs compute.
Shows a no-training route to fixing OCR failure modes, though it is an anecdotal, unverified report.
- OCR and document AIProject
MiniCPM-OCR 1.0 and Jev: open OCR model and output checker
A post describing MiniCPM-OCR 1.0, a 1.3B document-reading model, and MiniCPM-OCR Jev, a 2B model that scores which regions of an OCR output may need human review. Reported benchmark figures are the author's citations.
Relevant to engineers building OCR pipelines who want to flag likely errors before text reaches RAG or agents.
- OCR and document AIProject
Unlimited-OCR-2 benchmarked on multi-page PDFs across Mac and DGX Spark
An engineer ran Baidu's open 3.3B document-parsing model on the same 6-page PDF across three machines, checking output against the PDF's text. The post reports speed figures and finds base mode misreads a page once requests reach 4 pages, while Gundam mode reads it correctly.
It gives concrete throughput and accuracy-checking evidence for running long-document OCR locally versus on a GPU server.
- OCR and document AIProject
MiniCPM-OCR suite: open-source OCR models and an accuracy checker
OpenBMB announces two open-source models: MiniCPM-OCR 1.0 for document parsing at 1B scale, and MiniCPM-OCR Jev, a 2B model that checks parsing accuracy for text, formulas, and tables from different OCR parsers. The author reports benchmark scores and throughput claims.
Offers an open model that verifies other OCR parsers' output, relevant to checking document extraction results.
- 16 days ago @MaziyarPanahi
Gemma 4 31B reads a handwritten prescription in an OCR test
A post reports that Google's Gemma 4 31B model read a handwritten doctor's prescription image better than earlier OCR models, in the author's own handwriting test. No link or benchmark details are given.
It shows a practical handwriting OCR check, but the single informal example is anecdotal and needs proper evaluation.
- 16 days ago @ivanfioravanti
Baidu releases open-source 3.3B OCR model for long documents
Baidu announced Unlimited OCR 2, an open-source 3.3B model that parses documents of any length with near-constant decoding latency and memory, per the author's claim of state-of-the-art results on MPDocBench and MarathonDocBench. The post says the license is MIT.
Shows an approach to long-document OCR with bounded memory, useful if you process multi-page PDFs or books.
- 16 days ago @Baidu_Inc
Baidu releases open-source Unlimited OCR 2 for long documents
Baidu announces Unlimited OCR 2, an open-source document parsing model. It uses a new SR-SWA mechanism to condense text-visual attention into fixed-size working references, claiming near-constant decoding latency and memory as documents grow.
Shows a technique for keeping OCR attention and memory bounded on long, multi-page documents.
- OCR and document AIProject
Open-source classifier for tax document pages
The GitHub project is a tax document page classifier built on Jev decisions. Its description claims 100% strict accuracy across 261 IRS forms at about $0.001 per page.
Engineers can inspect an open-source approach to classifying tax document pages and its reported accuracy and cost.
- 1 month ago @NielsRogge
WeVisDoc and jina-ocr-v1 are new OCR models
The post reports Tencent’s WeVisDoc 2B and 4B models under Apache 2.0 and jinaai’s jina-ocr-v1 under a non-commercial license. It says WeVisDoc-4B scores 95.38 on OmniDocBench v1.6, while NaviDC-OCR is current SOTA.
The model options, license terms, and benchmark result help engineers compare OCR candidates.
- OCR and document AIProject
iOS OCR server using Apple’s Vision framework
The open-source iOS-OCR-Server project uses Apple’s Vision framework to provide OCR from an iPhone. The post describes local processing and an API for use by devices on the same network.
It offers engineers a way to explore on-device OCR through a network-accessible API.
- OCR and document AIProject
Testing Unlimited OCR Against an OCR Benchmark
The author says Baidu Unlimited OCR was added to mlx-vlm and plans to test it against an OCR benchmark. The post also asks for other strong OCR benchmarks.
Engineers evaluating OCR models can use the linked benchmark as a starting point.
- OCR and document AIProject
Unlimited OCR for Long Documents
Baidu’s open-source OCR model uses Reference Sliding Window Attention to process documents longer than 40 pages in one forward pass. The post reports 3B total parameters, 500M activated, and results on OmniDocBench v1.5 and v1.6.
Engineers can inspect the implementation of an approach to long-document OCR with constant KV Cache size.
- OCR and document AIProject
PaddleOCR 3.7 adds ONNX Runtime support
PaddleOCR 3.7 adds an ONNX Runtime inference backend and support for CPU inference via OpenVINO and GPU inference via CUDA and TensorRT. The post says PP-OCRv6 models are up to 3.9× faster on CPU.
Engineers can compare inference backends and hardware options when deploying OCR across platforms.
- OCR and document AIProject
PaddleOCR releases PP-OCRv6 model series
PP-OCRv6 includes Tiny, Small, and Medium models ranging from 1.5M to 34.5M parameters. PaddlePaddle reports support for 50 languages, new OCR scenarios, and up to 5.2× faster CPU inference with OpenVINO.
Engineers can compare model sizes, language coverage, and inference options for OCR deployments.
- 5 months ago @AdinaYakup
Baidu releases PP-OCRv6 for multilingual OCR
The post says PP-OCRv6 comes in 1.5M, 7.7M, and 34.5M parameter sizes, supports 48+ languages and several text types, and is designed for edge deployment.
The model-size options and stated text coverage may help engineers assess it for OCR workloads.
- 5 months ago @Baidu_ERNIE
PaddleOCR-VL 1.6 reports 96.33% on OmniDocBench v1.6
The post announces PaddleOCR-VL 1.6 and reports a 96.33% score on OmniDocBench v1.6, which it describes as a new state of the art.
The benchmark result may help engineers track OCR model performance on document understanding.
- OCR and document AIModel
GLM-OCR: a 0.9B multimodal OCR model
The post announces GLM-OCR, described as a 0.9B model using a multimodal GLM-V architecture and an MIT license. The author claims it ranks first on OmniDocBench v1.5 with a score of 94.62.
Engineers evaluating OCR models can inspect its model page and compare the reported benchmark claim.
- OCR and document AIArticle
Ente’s on-device OCR for Android
Ente describes how it brought private, on-device text recognition to Ente Photos. The post says its OCR is fully open source.
Useful as an implementation example for engineers exploring private, on-device OCR.
- OCR and document AIModel
Tencent releases the 1B-parameter Hunyuan OCR model
Tencent’s end-to-end document-understanding model supports document parsing, visual question answering, and translation. The post links to its Hugging Face model and an interactive demo.
Engineers can evaluate a document AI model across parsing, visual Q&A, and translation tasks.
