An open OCR model that tops benchmarks and preserves page layout

What it does

Chandra OCR 2 ingests images and PDFs and emits structured HTML, Markdown, or JSON while preserving the visual logic of the original page—tables, math, forms, checkboxes, and multi-column layouts. It handles more than 90 languages and can run locally via HuggingFace or remotely through a vLLM server. The project is backed by a managed API tier, though the open weights carry a modified OpenRAIL-M license that favors research, personal projects, and early-stage startups under $2M revenue.

The interesting bit

The maintainers built a 90-language benchmark from scratch because public OCR test sets are largely English-centric. On their tables, the open model outperforms olmOCR 2, GPT-4o, and Gemini Flash 2 on math, old scans, and complex tables—suggesting the architecture actually understands document geometry rather than simply reading text left-to-right.

Key highlights

• Ranks above olmOCR 2, GPT-4o, and Gemini Flash 2 on the project’s olmocr benchmark table, and shows strong gains on an internal 90-language suite.
• Reconstructs complex layouts: tables, math, forms with checkboxes, images with captions, and multi-column text.
• Two inference modes: a local hf backend or a remote vLLM server for heavier batch work.
• Every processed file yields parallel Markdown, HTML, and JSON outputs alongside metadata and extracted images.
• OpenRAIL-M license allows free research and personal use; startups under $2M revenue are covered, but commercial self-hosting requires a separate license and competitive use is barred.

Caveats

• The hosted Datalab API still edges out the open Chandra 2 weights on the project’s own benchmark table, so self-hosters sacrifice a few accuracy points for privacy.
• The license includes a non-compete clause: you cannot use the open weights to build a service that competes with the Datalab API.

Verdict

A solid choice for developers who need to turn scanned textbooks, handwritten notes, or financial forms into structured data without flattening them into plain text. Avoid it if you need a fully unrestricted commercial license or plan to launch a competing document-extraction API.