Xiaomi Open-Sources Xiaomi-OCR-0, a 0.8B Document Model Under Apache 2.0
The 0.8-billion-parameter vision-language model handles document parsing, key information extraction and OCR-based visual question answering in a single model, with weights and code released under a fully permissive license.
Xiaomi Research has published Xiaomi-OCR-0, a compact 0.8-billion-parameter vision-language model trained on roughly 170 million OCR-centric samples and released under the Apache 2.0 license, covering both weights and code. The model unifies three tasks: document parsing, where it reports an OmniDocBench v1.6 overall score of 96.83, key information extraction, and OCR-centric visual question answering, where it averages 83.2 across five benchmarks and outscores the larger Qwen3.5-2B. Outputs come as Markdown or JSON, and the model tolerates skewed scans, photos and curved text without pre-correction.
Deployment is comparatively light: GGUF and Ollama builds let it run on mid-range GPUs or CPUs, and the team ships an MCP service, a browser demo and a two-stage batch recipe built on PP-DocLayoutV3. Benchmark figures are self-reported, and third-party coverage notes that some Real5 scores depend on the team's own scoring code, so independent validation is still pending. Even so, the permissive license plus built-in KIE makes it an unusually enterprise-friendly entry in the current wave of sub-1B document parsers.