Reportedly: Xiaomi open-sources 0.8B Xiaomi-OCR-0 document understanding model
Model: Xiaomi-OCR-0 · availability, license and releases
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Xiaomi released and open-sourced Xiaomi-OCR-0, a 0.8B-parameter document understanding model combining document parsing, key information extraction, and visual question answering. It scored 96.83 on OmniDocBench v1.6, ranking first among sub-1B models. Weights, inference code, a browser demo, MCP service, and Agent Skill toolkit are publicly available, Zhidongxi reported.
China context
- Original name
- 小米
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and use the provided inference code, MCP service, and Agent Skill toolkit to build document understanding applications without relying on Xiaomi's cloud services.
- For investors
- Xiaomi's open-weight release of a competitive sub-1B document model signals its push into on-device AI, which could reduce its dependence on external model providers and strengthen its consumer device ecosystem.
Translated from Chinese. Quotes and facts link to the original sources.
Xiaomi has released Xiaomi-OCR-0, a lightweight 0.8B-parameter model that unifies document parsing, key information extraction, and OCR visual question answering. The model is designed for on-device use cases such as extracting takeout codes, calendar events from posters, and bill details. It achieves 96.83 on OmniDocBench v1.6, the highest among models under 1B parameters, and averages 83.2 on five document VQA benchmarks, outperforming HunyuanOCR, TokenVL, TextHawk, Qwen3.5-2B, and MiniCPM-V-4.5, while approaching the larger Qwen3.5-4B. Xiaomi has released the model weights on Hugging Face and the inference code on GitHub, along with a browser-based interactive demo, an MCP service, and an Agent Skill toolkit. The model powers the 'Super Xiao Ai' one-click screen recognition feature on Xiaomi phones.
Xiaomi-OCR-0 uses a 0.8B-parameter architecture that jointly models document parsing, key information extraction, and visual question answering. It supports structured output in Markdown and LaTeX for document parsing, JSON for key information extraction, and handles cross-region retrieval, numerical calculation, and semantic reasoning for VQA. The model is designed to handle complex layouts including dense formulas, nested tables, and distorted or noisy document images.
Xiaomi's release of a sub-1B document understanding model that outperforms larger rivals on OmniDocBench v1.6 puts pressure on other mobile AI providers to match on-device OCR and document parsing capabilities without relying on cloud APIs. Developers outside China can now integrate a competitive open-weight model for document workflows, reducing dependence on proprietary OCR services.
The model enables on-device document understanding for Xiaomi's Super Xiao Ai assistant, potentially reducing cloud processing costs and latency for screen recognition features. For developers, the open weights and tooling lower the barrier to building document automation applications.
A verifiable next signal is whether Xiaomi-OCR-0 appears in third-party benchmarks or is adopted in open-source document processing pipelines outside China. Another signal is whether Xiaomi extends the model to more languages or releases larger variants.