Indra OCR has public ONNX weights, a CLI and a Python package.
Datalab releases a 1M-parameter CPU OCR pipeline
2026-10-15Frontier watchTRIAL
Datalab released the Indra OCR models, CLI and Python package. Its main recognizer has about 1M parameters and the complete ONNX model bundle is about 18 MB. It converts PDFs and images on CPU into Markdown, text or boxed JSON, with layout, reading order and hybrid PDF text-layer handling. The authors report about 230 ms per page on four x86 vCPUs.
What this means for you
For local RAG and document-ingestion teams, a small CPU-only pipeline with structured layout output could reduce preprocessing cost. Throughput and accuracy are author-measured on specified data and hardware, while complex tables, handwriting, degraded scans and commercial license terms remain constraints.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Model | about 1M recognizer parameters; about 18 MB ONNX bundle |
|---|---|
| Outputs | Markdown, text or JSON with line and layout boxes |
| Author Performance | about 230 ms/page on four x86 vCPUs; up to 49 pages/s in the stated throughput setup |
| Limits | weak on complex tables, handwriting and degraded scans |
| License | Apache-2.0 code; modified OpenRAIL-M weights with commercial restrictions |
| Time Basis | HF weights committed 2026-10-15T06:40:56Z; found by recall audit |
Technical impact 2
Engineering value 4
Adoption signal 2
Maturity 3
Verification cost 2
Risk 3