Xiaomi-OCR-0: open 0.8B model reads complex documents

Xiaomi unveiled Xiaomi-OCR-0 on October 30, 2026, a vision-language model with just 0.8 billion parameters built to understand documents. The announcement came from the Xiaomi Tech account on X, together with model weights, inference code, an online demo, an MCP service and an Agent Skill.

The model handles three tasks in one package: parsing a page into Markdown, including tables and formulas; extracting specific fields as JSON; and answering questions about an image of text. It starts from the Qwen3.5-0.8B-Base model and was trained on a corpus of about 170 million OCR-focused examples.

What the reported scores say — and what is still unconfirmed

These scores are Xiaomi's own claims, not independent verification. The company reports 96.83 on OmniDocBench v1.6 and 95.24 on Real5-OmniDocBench, plus an average of 83.2 across five document question-answering tests. Among models under 1 billion parameters, Xiaomi says it leads document parsing on OmniDocBench v1.6.

One caveat matters: on the public Real5-OmniDocBench leaderboard on Hugging Face, the model scores 92.73, behind PaddleOCR-VL-1.6 at 93.19. The gap was attributed to the evaluation code version: Xiaomi used the v1.6 code, while the public leaderboard was aligned with v1.5. Real-world performance on a specific archive should therefore be tested before any critical use.

Open licence and local use

Weights and code are distributed under the Apache 2.0 licence, which allows commercial use under its terms. Local runs need runtimes such as SGLang, vLLM or Ollama, and with Transformers the model can also start on CPU. Xiaomi says the model already runs in its Super XiaoAi phone assistant.

Confidential documents should not go to the hosted Hugging Face Spaces demo, where the image leaves the user's machine. In a local run, processing can stay on the user's computer.

Sources

Image credit: illustration from the Xiaomi-OCR-0 article published by Pasquale Pillitteri.

Transparency: this article was produced with AI assistance, cross-checked against public posts on X and internet sources. Consult the original sources for full details.

Muse