Optical character recognition, the technology that turns scanned PDFs into machine readable text, just got a real open benchmark index alongside new Baidu and Mistral models in the same week, though the "open" in open source is doing more work than it should.
Open-source OCR just crossed a credibility threshold for production AI agent stacks. That sentence sounds like hype. It is not. In the same week, Baidu released Unlimited OCR, a 3-billion-parameter model that the company says can transcribe 40-plus-page documents in a single pass without losing context, and Mistral launched OCR 4 as a structured-output document intelligence product aimed at retrieval and agent pipelines. Meanwhile, the long-running machine-learning index Papers with Code now hosts a consolidated OCR task page mapping the major benchmarks to their current top open models, with links to papers, code, and weights. The combination is genuinely new. There is finally enough open OCR, and enough open OCR benchmarking, to choose a model on something other than vibes.
What OCR is, and why the timing matters: optical character recognition is the technology that turns scanned PDFs, photographed receipts, and ancient TIFFs into clean text an AI system can read. For most of the last decade, OCR meant either an expensive commercial API (ABBYY, Textract, Google Document AI) or a mediocre open model that choked on tables and handwriting. That tradeoff was acceptable when the downstream consumer was a human. It is not acceptable when the downstream consumer is a large language model that has to find a single line item in an 80-page contract. Modern AI agents and retrieval-augmented chatbots fail quietly on bad OCR in ways a human would have caught. A misread number becomes a confidently wrong answer.
That is the wedge Baidu is pushing into. According to the arXiv preprint, Unlimited OCR builds on earlier open OCR work, specifically DeepSeek OCR, and introduces a mechanism the company calls Reference Sliding Window Attention, an attention pattern that lets the model keep long-range context across dense pages without paying the quadratic cost of full self-attention. The model is being positioned as a drop-in alternative for the messy, mixed-layout documents that wreck vanilla OCR. Weights are live on Hugging Face, code is on GitHub, and Baidu's official announcement leans hard on the "3B total parameters and 500M activated" framing alongside the one-pass, 40-plus-pages claim. Treat that 40-page number as a released claim, not a verified one. The preprint documents the experimental setup, but the headline result has not been independently reproduced by a third party as of late June 2026.
Mistral is taking a different route. OCR 4 is an API product, not open weights, and Mistral is selling it as the document-understanding layer for RAG and agentic systems that need structured output, not just raw text. Coverage in MarkTechPost frames it as a competitor to the closed-API incumbents rather than to the open-weight crowd, with further recaps and developer explainers echoing that positioning. That distinction matters, because the "open-source OCR is finally good" framing routinely elides it. An API you cannot self-host is not, in any operational sense, an open-source replacement for a commercial OCR service. It is just a cheaper commercial OCR service.
The third piece of the story is the Papers with Code OCR index itself, which for the first time gives a curious engineer a single page that maps the major open benchmarks, with the maintainer highlighting OlmOCRBench from Ai2 and OmniDocBench from the Shanghai AI Laboratory, to the current top open models. The maintainer's flagged picks lean toward Chandra OCR 2 and Mistral OCR v4. That is useful as a shortlist, not as a verdict. The maintainer is doing curation, not running head-to-head evaluations, and his "top" reflects his judgment. The "best open-source OCR" question still requires the reader to do the actual evaluation on their own documents, in their own pipeline, against their own failure modes.
What it means in practice: if you are choosing OCR for a production agent stack this week, the honest answer is that the open-weight side just got a serious new candidate in Baidu Unlimited OCR, worth piloting on long, dense, mixed-layout PDFs before you sign a multi-year API contract. The closed side has a serious new candidate in Mistral OCR 4, worth pricing against whatever you are using now. And for the first time, there is a single index where the next person to release a 3-billion-parameter OCR model will, presumably, get ranked. The next model worth piloting is probably already on a Hugging Face repo this week. Whether the index can keep up with the release cadence is the open question.