Datalab's OCR Lite: Launch and Benchmark Domination

Episode 1 · Datalab Launches $0.5/1K-page OCR Lite Mode and Open-Sources 1M-parameter Indra (2026-10-22, 8 posts)

Datalab founder Vik Paruchuri rolled out new OCR offerings on October 22–23: a lite mode priced at $0.50 per 1,000 pages, along with the open-source release of its underlying model Indra (just 1 million parameters, runs purely on CPU). The lite mode delivers full document structure output at an extremely low price, outperforming Azure's comparable products in benchmarks—with an especially wide gap on non-Latin scripts. Indra is available on Hugging Face and can be deployed without PyTorch or a GPU.

Confirmed

• Pricing: lite mode costs $0.50 per 1,000 pages, forming a three-tier lineup alongside balanced ($4/1K pages) and accurate ($10/1K pages); lite is 8x cheaper than their own balanced tier
• Capabilities: lite returns full document structure including text, tables, headings, lists, reading order, and layout boxes for every block; long documents are processed 2x faster than cloud vendors
• Model: Indra has just 1M parameters, runs on CPU via ONNX Runtime with no PyTorch/GPU needed, at 44 pages/second, supporting 80+ languages; balanced/accurate modes use larger models
• Benchmarks: on OmniParseBench's no-math subset (12,972 test items), lite scores 75.1 overall vs. just 42.3 for Azure Read; its text score also beats Azure Layout; the Indra model itself scores 66.7 on OmniParseBench (71.4 with --math enabled)
• Multilingual edge: results on non-Latin scripts are lopsided—for Armenian, lite hits 97% while Azure manages only 21%, with similar leads on languages like Ethiopic
• Cost comparison: 20–30x cheaper than Azure, Google, and AWS layout APIs; roughly twice as accurate as Azure Read at half the price, and 10x cheaper than Azure Layout

Limitations

• Paruchuri himself notes that lite struggles with handwriting and complex layouts (including tables), so those use cases still require the pricier balanced/accurate modes

Why it matters

• A tiny 1M-parameter model running on CPU achieves OCR quality close to much larger models, dramatically lowering the deployment barrier and cost of document parsing—a direct draw for budget-sensitive, high-throughput scenarios like bulk document digitization and multilingual document processing; pairing an open-source model with a low-price API also puts pricing pressure on cloud vendors.

• Datalab's Indra OCR: a 1M-parameter open model that hits 44 pages/s on CPUs, beating PaddleOCR — generativist · 2026-10-22
• Datalab open-sources Indra OCR: a 1M-parameter model that beats PaddleOCR on CPU — VikParuchuri · 2026-10-23
• Datalab lite doubles Azure Read accuracy at half the price on OmniParseBench — VikParuchuri · 2026-10-23
• Datalab launches lite OCR at $0.50 per 1,000 pages, 20-30x cheaper than Azure, Google and AWS layout APIs — VikParuchuri · 2026-10-23
• Datalab lite: full-layout OCR at $0.50/1,000 pages, more accurate than Azure Read at half the price — VikParuchuri · 2026-10-23
• Datalab OCR tiers: lite at $0.50, balanced at $4.00, accurate at $10.00 per 1,000 pages — VikParuchuri · 2026-10-23
• Datalab's Indra OCR: a 1M-parameter open model that hits 44 pages/s on CPU across 80+ languages — VikParuchuri · 2026-10-23
• Datalab lite OCR scores 97% on Armenian vs Azure Layout's 21%, benchmark shows — VikParuchuri · 2026-10-23

Episode 2 · Marker lite mode beats Azure Read in document parsing benchmark (2026-10-23, 2 posts)

Marker's lite mode scored 75.1 on OmniParseBench (documents without math formulas), far ahead of Azure Read at 42.3, with strong accuracy on non-Latin scripts, though the author acknowledged its limitations.

• Doc parsing lite mode scores 75.1 vs Azure Read's 42.3 on OmniParseBench, dominates non-Latin scripts — VikParuchuri · 2026-10-23
• Marker's lite mode scores 75.1 vs 42.3 for Azure Read on OmniParseBench — VikParuchuri · 2026-10-23