Xiaomi Open-Sources 0.8B Document Understanding Model Xiaomi-OCR-0
Xiaomi has released and open-sourced a lightweight document understanding model, Xiaomi-OCR-0, with only 0.8B parameters, integrating three capabilities—document parsing, key information extraction, and OCR visual question answering—into one. The model scored 95.24 on Real5-OmniDocBench, ranking first among compared models, and scored 96.83 on OmniDocBench v1.6, ranking first among models under 1B parameters, with an average score of 83.2 across five mainstream OCR visual question answering benchmarks. Xiaomi has also open-sourced the model weights and inference code, and provides a Demo, MCP service, and Agent Skill toolkit. The related technology has already been used in Xiaomi phones and "Super Xiao Ai" for screen information recognition such as pickup codes, bills, and cards.
Why read this
AI score 63
Xiaomi has open-sourced a 0.8B document understanding model, providing multiple benchmark scores and on-device smartphone deployment scenarios, allowing comparison of the parameter-efficiency ratio of lightweight models on OCR tasks.
Topics
Body · AI translation
Xiaomi Open-Sources 0.8B "King of Competition" OCR Model: 800 Million Parameters, Punching Above Its Weight Against Large Models
Phoenix Tech Download App | Exclusive First Look October 31, 2026, 11:51:57 | From Beijing | Updated October 31, 2026, 11:52:28
Phoenix Tech News (Author/Yu Lei) — On October 31, Xiaomi yesterday released and open-sourced Xiaomi-OCR-0, a lightweight document understanding model. With only 0.8B parameters, the model integrates three capabilities—document parsing, key information extraction, and OCR visual question answering—into one, aiming to achieve structured parsing and semantic understanding of complex documents at a lightweight scale.
In terms of benchmark performance, Xiaomi-OCR-0 scored 95.24 on Real5-OmniDocBench, a benchmark of real-world photographic interference scenarios, ranking first among the compared models.
On OmniDocBench v1.6, it scored 96.83, taking second place overall and first among models under 1B parameters with its 0.8B parameters.
On five mainstream OCR visual question answering benchmarks, it achieved an average score of 83.2, surpassing some dedicated OCR models and larger general-purpose multimodal models, demonstrating a high parameter efficiency ratio.
On the technical side, Xiaomi says the model was trained on approximately 170 million high-quality samples. Its core lies in a scalable data engine and a progressive training strategy of "parse first, understand later," which reduces pseudo-label noise and fills in capabilities for complex layouts and long-tail scenarios.
In terms of deployment, the related technology has already entered Xiaomi phones and the "Super Xiao Ai" scenarios, where it can be used to recognize and subsequently act on on-screen information such as meal pickup codes, bills, cards and certificates, schedules, and flight information. Xiaomi has also open-sourced the model weights and inference code, and provides a Demo, MCP service, and Agent Skill toolkit for developers to integrate.