Datalab's Chandra OCR 2 tops benchmarks

Chandra 2 reads like Datalab trying to own the structured document intelligence layer, not just OCR. If the benchmark claims hold up outside the vendor eval, the mix of layout-aware decoding, handwriting support, and flexible deployment makes it a compelling default for serious document pipelines.

• –The full-page decoder is the right architectural call for messy docs: handwriting in forms, cross-row tables, and image extraction are exactly where pipeline OCR tends to break.
• –Datalab's own 90-language benchmark says Chandra 2 averages 72.7% vs Gemini 2.5 Flash at 60.8%, which is a real claim but still one the market should pressure-test on real workloads.
• –The deployment ladder is strong: local Hugging Face, remote vLLM, free playground, and hosted API make it easy to start open and scale later.
• –The licensing split matters: the code is Apache 2.0, but the model weights use modified OpenRAIL-M terms, so commercial users need to read the fine print before treating it as fully permissive open source.