Xiaomi-OCR-0: A 0.8B Open Model for Parsing, Extraction and OCR Q&A

OCR has suddenly become a crowded field. Big tech and startups alike are shipping small document models, competing to take a PDF page apart cleanly with as few parameters as possible. Xiaomi's entry, Xiaomi-OCR-0, has just 0.8B parameters and tries to handle three jobs in one: document parsing, key information extraction (KIE) and text-centric visual question answering.

The scope is well defined. Parsing returns structured Markdown, with tables generated as OTSL and converted to HTML tables, and formulas in LaTeX. KIE returns JSON for the fields you specify. VQA answers questions about the text and content of a document image. The model starts from Qwen3.5-0.8B and is trained on about 170 million OCR-centric samples in three stages: Q-Mask text anchoring to align fine-grained text with its position, multi-task continued pretraining, then mixed-task reinforcement learning with verifiable rewards on hard examples.

The scorecard is strong. It ranks first on Real5-OmniDocBench at 95.24 and second on both OmniDocBench v1.6 (96.83) and Wild-OmniDocBench (87.94). It scores 97.84 on UniMER-Test for formulas and 92.27 on PubTabNet for tables, and averages 83.2 across five OCR VQA benchmarks, beating Qwen3.5-2B, a model more than twice its size. On historical Chinese handwriting and calligraphy, it ties Seed 2.0 Pro.

The data engine deserves more attention than the scores. Rather than leaning on large-scale human labeling, Xiaomi has several heterogeneous expert models vote toward consensus, verifies labels by rendering them back and comparing, and adds hard-sample mining and targeted synthesis. That recipe is what lets a small model keep improving, and it is the most reusable part of the technical report.

Why 0.8B? Look at Xiaomi's business. Receipts, screenshots and IDs in the phone gallery, photographed pages in notes apps, manuals in the car cockpit: all of these want recognition on-device, which saves cloud cost and protects privacy. Xiaomi documents SGLang and vLLM deployment, and the community has already shipped GGUF quantizations and an Ollama package. Quantized, it fits in under 1GB and runs on an ordinary laptop.

Small-model weaknesses remain. Community tests have caught look-alike character mix-ups, such as a comma in an amount read as a period, so financial fields still need a second check. But released under Apache-2.0 and cleared for commercial use, Xiaomi-OCR-0 could easily become the default first step in many RAG and document-agent pipelines.