OXR
An open-source document parsing model developed by the SoMarkAI team with 1.3B parameters. It converts PDFs and document images into Markdown or JSON with layout coordinates, recognizes 13 element types, and employs pluggable layout analysis with a recursive self-improvement post-training loop.
Timeline (last 90 days)
OXR is an open-weight document parsing model with only 1.3B parameters, developed by the SoMarkAI team, aiming to convert PDFs and document images directly into Markdown or JSON with layout coordinates.
OXR officially claims to rank first simultaneously on three mainstream document parsing benchmarks: olmOCR-Bench, OmniDocBench v1.6, and DocParsingBench.
In a test running all 21 models on a fixed set of 100 OmniDocBench pages on a single RTX PRO 6000, OXR end-to-end (including layout analysis, recognition, and reading order) processed 100 pages in 101 seconds, while the next fastest, OvisOCR2, took 135 seconds.
OXR supports recognizing 13 element types and outputting bounding boxes, including tables, formulas, chemical equations, chemical structure diagrams, and code blocks.
OXR's engineering design includes an RSI post-training loop, which uses recursive self-improvement to automatically convert parsing errors and uncertain results into targeted training data without manual annotation.
OXR decouples layout analysis from recognition: recognition and reading order are executed in parallel within the same model to ensure speed, while layout analysis is an independent model that can be replaced as needed.
Running OXR locally requires Linux and an NVIDIA GPU with BF16 support; 24GB of VRAM is enough for low-concurrency testing, while long documents or high concurrency call for 48GB or more; the Docker all-in-one image is the easiest way to get started.
OXR offers a free API that requires no registration or API key, but it is for testing only and requests are queued; the team said it will stay online until it can no longer afford the cost.
OXR's repository is at GitHub's SoMarkAI/OXR, the weights are at Hugging Face's SoMarkAI/OXR-1.0, and it is released simultaneously on ModelScope; the license is Apache 2.0 with additional terms, including threshold restrictions on commercial use.
All Facts (9)
OXR is an open-weight document parsing model with only 1.3B parameters, developed by the SoMarkAI team, aiming to convert PDFs and document images directly into Markdown or JSON with layout coordinates.
50%UnverifiedOXR officially claims to rank first simultaneously on three mainstream document parsing benchmarks: olmOCR-Bench, OmniDocBench v1.6, and DocParsingBench.
50%UnverifiedIn a test running all 21 models on a fixed set of 100 OmniDocBench pages on a single RTX PRO 6000, OXR end-to-end (including layout analysis, recognition, and reading order) processed 100 pages in 101 seconds, while the next fastest, OvisOCR2, took 135 seconds.
50%UnverifiedOXR supports recognizing 13 element types and outputting bounding boxes, including tables, formulas, chemical equations, chemical structure diagrams, and code blocks.
50%UnverifiedOXR's engineering design includes an RSI post-training loop, which uses recursive self-improvement to automatically convert parsing errors and uncertain results into targeted training data without manual annotation.
50%UnverifiedOXR decouples layout analysis from recognition: recognition and reading order are executed in parallel within the same model to ensure speed, while layout analysis is an independent model that can be replaced as needed.
50%UnverifiedRunning OXR locally requires Linux and an NVIDIA GPU with BF16 support; 24GB of VRAM is enough for low-concurrency testing, while long documents or high concurrency call for 48GB or more; the Docker all-in-one image is the easiest way to get started.
50%UnverifiedOXR offers a free API that requires no registration or API key, but it is for testing only and requests are queued; the team said it will stay online until it can no longer afford the cost.
50%UnverifiedOXR's repository is at GitHub's SoMarkAI/OXR, the weights are at Hugging Face's SoMarkAI/OXR-1.0, and it is released simultaneously on ModelScope; the license is Apache 2.0 with additional terms, including threshold restrictions on commercial use.
50%