Patent-Specialized 0.3B OCR Model MOSS-OCR Released

MOSS-OCR, a model designed to handle the complex structures unique to patent documents (merged tables, chemical structure formulas, mixed CJK/Latin formulas, etc.), has been released. This model excludes the layout detection stage and focuses on block-level text recognition, maintaining high efficiency with a small parameter size of 0.3B.

Key performance and features are as follows:

• Performance: On OmniDocBench v1.6, it achieves a score similar to MinerU 2.5 (94.46) while using only about 25% of the parameters. On its own patent benchmark (patent-bench), it ranked 1st (93.49).
• Speed: On a single RTX 4090, it delivers a throughput of about 59 QPS via vLLM, roughly 2x faster than MinerU 2.5.
• Configuration: For cases requiring layout detection, it also provides the Hiro-Smart-Doc pipeline, which combines an RT-DETR-based layout detection model with MOSS-OCR.

All code has been released under the Apache 2.0 license.