RIP Commercial OCR. An Open-Source Model Just Topped Every Benchmark.
Datalab’s 4 billion-parameter Chandra OCR 2 outperforms GPT-4o and Gemini across independent and multilingual benchmarks, handling complex layouts, math notation, flowcharts and 90 languages with state-of-the-art accuracy. It’s available under Apache 2.0 code with a modified OpenRAIL-M license for weights, runs locally via HuggingFace or vLLM, and doubles throughput over its predecessor.
- Chandra OCR 2 (4B params, open-weight) scored 85.9% on olmOCR vs GPT-4o's 69.9%, and beat Gemini/GPT-5 Mini on multilingual benchmarks, with huge gains on South Asian scripts (Kannada +42.6, Malayalam +46.2, Telugu +39.1).
- It processes full pages in one pass rather than splitting into blocks, giving it an edge on tables, nested headers, checkboxes, handwritten math, and flowcharts exported as Mermaid diagrams.
- Despite shrinking from 9B to 4B parameters, throughput doubled to ~2 pages/sec on an H100 while accuracy improved.
- Code is Apache 2.0 and installable via pip/Docker, but weights use a modified OpenRAIL-M license requiring a paid commercial license for larger companies.
Datalab’s Chandra OCR 2 is a 4-billion-parameter open-weight model that converts images and PDFs into structured Markdown, HTML or JSON while keeping original layouts intact. On the independent olmOCR benchmark from AllenAI, it scored 85.9%, beating GPT-4o’s 69.9%. In Datalab’s own 90-language test, Chandra OCR 2 hit 72.7% versus Gemini 2.5 Flash at 60.8%. For the 43 most common languages it climbed to 77.8%, compared with GPT-5 Mini’s 60.5%. South Asian scripts saw the biggest jumps—Kannada up 42.6 points, Malayalam 46.2, Telugu 39.1—making a real difference if you work with those writing systems.
Unlike pipelines that break pages into blocks, Chandra processes the full page in one pass, which explains its edge on tables, nested headers, forms with checkboxes, handwritten math and even flowcharts output as Mermaid diagrams. Paruchuri’s team cut the model size from 9B to 4B parameters, doubled throughput to about two pages per second on an H100 GPU, and still improved accuracy. You can run it locally via pip install chandra-ocr[hf] or deploy with Docker. The code is Apache 2.0, but the weights use a modified OpenRAIL-M license—free for research and small startups, commercial license required for larger companies. Third-party multilingual tests and careful checks on handwriting-form extraction are recommended before you build a production pipeline around it.