Datalab's Indra OCR 1.1 Brings Handwriting and 80+ Language OCR to CPU-Only Deployments
The ~1M-parameter recognizer runs entirely on CPU, claiming 11-17x speedups over Tesseract and PaddleOCR PP-OCR.
Datalab, the maker of Marker and Surya, released Indra OCR 1.1, an open-source OCR pipeline whose recognizer has roughly 1 million parameters, with the default pipeline around 7M parameters and 18MB, running on ONNX Runtime without PyTorch or a GPU. The vendor claims 49 pages per second on 32 CPU cores, about 230ms per page on 4 vCPU, and 11-17x speedups over Tesseract and PaddleOCR's PP-OCR. Version 1.1 adds handwriting recognition, scanning for over 80 languages including Arabic, Hebrew, Thai and Indic scripts, and page straightening for skewed scans.
The company reports an OmniParseBench score of 66.7, or 71.4 with math enabled, and is candid that Indonesian is not yet among the publicly tested scan languages and that complex tables remain weak, with harder documents referred to its Chandra model. The code is Apache 2.0 while the weights are under OpenRAIL-M, which carries usage restrictions. Benchmark figures are self-reported via Datalab's own harness.