OpenBMB has released the first specialized OCR model suite, MiniCPM-OCR: the base model with 1.3 billion parameters recognizes text, tables, formulas, and stamps, while a separate MiniCPM-OCR-Jev model with 2 billion parameters automatically evaluates recognition accuracy and flags suspicious areas for manual review. In parallel, Baidu released Unlimited-OCR 2 with approximately 3 billion parameters: the SR-SWA mechanism compresses attention to 512 anchor tokens, so the cost of a decoding step barely grows with document length. Both models are open-sourced along with code and server stacks, allowing a local PDF parsing pipeline with built-in quality checks to be assembled today.

image
image
image

What happened

OpenBMB published the MiniCPM-OCR collection on Hugging Face. The base MiniCPM-OCR 1.0 is built on MiniCPM-V 4.6 (SigLIP and Qwen3.5) and contains 1.3 billion parameters; it recognizes text, tables, formulas, and stamps and, according to developers, scores 93.68 on Real5-OmniDocBench. DSpark acceleration increases throughput by 55%, reaching 1.2 pages per second on an RTX 5090; speed is claimed to be 1.49 times higher than PaddleOCR-VL-1.6 and 3.62 times higher than MinerU2.5 Pro. The separate MiniCPM-OCR-Jev model with 2 billion parameters, based on Qwen3.5-2B, checks the accuracy of OCR results, including output from third-party parsers, and achieves 0.936 AUROC on JevOCRBench. Baidu released Unlimited-OCR 2 with approximately 3 billion parameters under the MIT license on October 27–28, 2026: in 100-page tests, individual decoding steps run up to 7 times faster than full attention and up to 5.6 times faster than Unlimited-OCR 1.0.

Context

Both releases address long-standing problems in document AI. The first is quality control: recognized text always contains errors, and industrial pipelines spend significant human time verifying disputed areas, so OpenBMB isolated this task into a separate judge model; the AUROC metric is methodologically appropriate as it is a threshold-independent measure suitable for separating reliable and suspicious fragments. The second problem is the quadratic growth of attention cost: the longer the document, the more expensive each decoding step becomes for classic transformers, which is why multi-page documents are usually processed in parts, losing cross-page context. The SR-SWA mechanism solves this by compressing attention to a fixed set of anchor tokens, keeping the step cost nearly constant. Finally, scale matters: models with 1.3–3 billion parameters are compact enough to run on local hardware, making these releases oriented toward real deployment rather than showcase demonstrations.

Why this matters for the industry

For the industry, both releases primarily change the economics of document AI. The class of judge models makes quality control partially automatic: reliable fragments are accepted by the pipeline itself, and only the suspicious remainder goes to humans—this directly reduces control costs in industrial OCR pipelines. SR-SWA eliminates the quadratic growth of attention cost on multi-page documents, making one-off parsing of hundreds of pages economically viable without losing cross-page context, and long documents cease to be a separate expensive case. Together, this brings document recognition closer to commoditization: open small models provide benchmark-level quality, document parsing becomes a standalone product available to startups and small teams, and value shifts to trust in extracted data, domain-specific fine-tuning, and integration with agentic processes. If independent reproductions confirm the claimed metrics, the "parser plus quality judge" pattern will likely become a standard block in industrial pipelines, similar judges will appear from other developers, and attention compression schemes like SR-SWA will likely be transferred to other long-context tasks for multimodal models.

Why this matters for users

Practically significant is that all components are available today. MiniCPM-OCR is available on Hugging Face in BF16, GGUF, AWQ, and GPTQ formats and works with vLLM, SGLang, llama.cpp, and Ollama, while Unlimited-OCR 2 has published code, an SGLang server, and an interactive demo on Hugging Face Spaces. This allows deploying local parsing of PDFs and scans with tables and formulas on a single consumer GPU and connecting Jev as an automatic quality check layer before recognized text proceeds to further processing. A reasonable first step is not to immediately replace the current parser, but to run a shadow mode: process a real stream of documents through both models and the judge, compare results on your own data, and only then choose the stack for production.

What is still unknown / limitations

All key figures for both releases—results on Real5-OmniDocBench and JevOCRBench, DSpark and SR-SWA accelerations, performance comparisons with PaddleOCR-VL-1.6, MinerU2.5 Pro, and Unlimited-OCR 1.0—are provided by developers in release materials, with no independent verification in the sources. Evaluation methodologies, including testing protocols, versions of compared models, and scientific publications, are not disclosed in the materials, so the correctness of the claimed comparisons cannot yet be verified. Since the weights of both models are open, independent reproductions should be expected in the near future; until then, it is reasonable to treat the claimed metrics as vendor data.

Sources

Author

Look at AI, editorial team