Chinese OCR models for document parsing

Twelve Chinese OCR models and Baidu's classic PaddleOCR pipeline, compared for full-page parsing precision, speed, hosted price, the boxes they return, and the input they take.

Key takeaways

• TeleOCR leads OmniDocBench v1.6 at 96.91, with OvisOCR2 (96.47) and PaddleOCR-VL-1.6 (96.34) within half a point.
• GLM-OCR and DeepSeek-OCR 2 have the lowest hosted price, $0.03 per million tokens; GLM-OCR scores 95.22 against 90.25.
• Only Baidu's classic PaddleOCR pipeline returns word, character and table-cell boxes; every vision-language parser stops at text lines or layout blocks.

The scoreboard

How it was measured

Precision is OmniDocBench v1.6's overall score for full-page parsing (text, tables, formulas and reading order), out of 100: the official leaderboard where a model is on it, otherwise the lab's own number, marked self-reported. Speed and price come from the party named in each row. Labs measure speed on their own hardware and concurrency, so speeds are rough order only. Compiled on Oct 2 and 3 2026 from public model cards, papers and API docs.

Log