Xiaomi Open-Sources a 0.8B Document Model That Beats 8B Rivals on OCR Benchmarks

Xiaomi has open-sourced Xiaomi-OCR-0, a document-understanding model that packs document parsing, key-information extraction and OCR-based question answering into a single 0.8-billion-parameter model released under Apache 2.0. Weights and inference code are on Hugging Face and GitHub, alongside a browser demo, an MCP service and an Agent Skill toolkit for wiring the model into agent workflows.

On paper, the small model punches well above its size. Xiaomi reports 95.24 points on Real5-OmniDocBench, a benchmark built around five realistic capture conditions including curved paper, screen rephotography and tilted shots, where it ranks first in its comparison table. On the broader OmniDocBench v1.6 it scores 96.83, second overall and just 0.04 points behind the 1.2-billion-parameter TeleOCR, and first among models under 1B parameters. Across five OCR question-answering benchmarks it averages 83.2, ahead of the 8-billion-parameter MiniCPM-V-4.5 at 82.6 and the 2-billion-parameter Qwen3.5-2B at 80.9.

Those numbers are the company's own and have not been independently verified, though the model card lists the comparisons: OvisOCR2 at 0.8B scores 96.58 on the same OmniDocBench table, PaddleOCR-VL-1.6 at 0.9B scores 96.33, and GLM-OCR at 0.9B scores 95.22. Formula recognition on UniMER-Test reaches 97.84 and table structure recognition on PubTabNet 92.27. On classical texts and calligraphy, the Chronicles-OCR benchmark, it scores 73, in the same range as the closed-source Seed 2.0 Pro at 72.

The model is built on Qwen3.5-0.8B-Base and trained on roughly 170 million OCR and document samples in three stages: a Q-Mask anchoring stage that ties text to its position on the page, continued OCR pretraining on a multi-task data mix, and mixed-task reinforcement learning that uses structure edit distance, table matching and answer accuracy as verifiable rewards. Training data was filtered by multiple OCR expert models voting on labels, and table and formula outputs went through a predict-rerender-compare loop that re-renders candidate parses and compares them visually against the original page.

One nuance worth noting: the weights themselves are not new. The Hugging Face repository was created on September 29 and the weights were posted on September 30 together with a technical report, quietly accumulating about 15,000 downloads before the formal announcement on October 30 brought the launch wider attention. A community Ollama package comes in at roughly 1.8GB and has been verified on Apple Silicon with Ollama 0.35.0, where seven test samples matched token counts and five matched the original output verbatim.

Xiaomi says the technology already runs inside its phones and the Super XiaoAI assistant, powering pickup-code extraction, bill logging, ID-card field capture and automatic calendar creation. That is the strategic point of the release: while general-purpose labs compete on scale, the OCR race is being won at the other end — small, license-clean models that run on-device, output straight into structured workflows, and cost almost nothing to deploy.