Overview
Xiaomi-OCR-0 is an open document understanding model from Xiaomi, 0.8B parameters, built on Qwen3.5-0.8B-Base and trained on roughly 170 million OCR and document samples. The pitch is that one small checkpoint covers three jobs: turning a full page into structured output in reading order, pulling specified fields out of receipts as JSON, and answering questions about documents and charts.
The point is where it runs, more than how large it is. Weights and inference code are released under Apache 2.0 on Hugging Face, ModelScope and GitHub, and a community Ollama build in bf16 weighs about 1.8GB, which fits on a laptop. Xiaomi also ships a browser demo, an MCP service and an Agent Skill toolkit, so an agent that installs the skill gains four tools: image parsing, PDF parsing, field extraction and image Q&A.
The benchmarks focus on real capture conditions. On Real5-OmniDocBench, which simulates scanning, curled paper, screen photos, lighting changes and tilted shots, it scores 95.24 and tops Xiaomi's comparison table. On OmniDocBench v1.6 it scores 96.83, first among models under 1B parameters and 0.04 points behind the leader overall, with an average of 83.2 across five visual question answering benchmarks. The technical report appeared on September 28 and Xiaomi published the full developer entry point on October 30.
Key Features
- Full-page parsing to Markdown: Recognises body text, tables and formulas and returns them in reading order, with tables as HTML and formulas as LaTeX.
- Field extraction to JSON: Targets invoices, receipts, ID cards and business forms, locates requested fields and returns JSON that feeds accounting or expense systems directly.
- OCR visual question answering: Goes past transcription to cross-region lookup, numeric calculation and reasoning, such as asking for a payment date in a contract.
- Region and whole-page modes: Regular printed documents run through layout detection and parallel region recognition, while handwriting, historical text and irregular layouts use whole-page parsing to keep context.
- MCP service and Agent Skill: The repository ships a SKILL.md, an MCP service and an install script, giving an agent image parsing, PDF parsing, field extraction and image Q&A.
- Small enough for local use: The 0.8B weights ship in a roughly 1.8GB community bf16 build, with SGLang and vLLM serving support on Python 3.10 or newer.
Use Cases
- Finance teams drop in a stack of invoices and receipts, extract amount, merchant and date as JSON, then push them into expense workflows.
- Knowledge bases parse scans and PDFs into Markdown before chunking, keeping table structure intact for retrieval.
- On phones the model already powers screen recognition, pulling pickup codes, parcel numbers and queue tickets from whatever is on screen.
- Contracts and certificates go through field extraction so clauses and document numbers become structured records.
- Historical, handwritten and calligraphic material runs in whole-page mode for long-tail text extraction.
- Teams run a local copy for documents that should not leave the network.
Pros
- 0.8B parameters rank first under 1B on document parsing benchmarks, which is a striking parameter efficiency.
- Parsing, field extraction and visual question answering live in one model instead of three services.
- Output lands as Markdown, HTML tables, LaTeX and JSON, ready for the next step.
- Apache 2.0 weights and code remove licensing friction for commercial and further training use.
- An MCP service and Agent Skill ship with the repository, so agent integration is one install.
- Xiaomi has validated it on real phone capture, which suggests it holds up on photos and screen content.
Pricing
Open source under Apache 2.0. Weights are on Hugging Face and ModelScope, while inference code, browser demo, post-processing, evaluation pipeline, MCP service and Agent Skill live in the SeerRay-Lab/Xiaomi-OCR-0 repository on GitHub. Local deployment requires Python 3.10 or newer and supports SGLang or vLLM, and the community Ollama build is about 1.8GB in bf16. Region mode also needs PaddlePaddle, PaddleX and the PP-DocLayoutV3 layout model. A hosted API has no published pricing yet.
Summary
Xiaomi-OCR-0 bets that the document path can be covered by one small enough model. Handling receipts and contracts used to mean wiring layout detection, table recognition, formula recognition and field extraction as separate services, where the plumbing often cost more than the model. Compressing those stages into one 0.8B checkpoint with unified Markdown and JSON output brings integration down to installing a package, which for scanned documents that cannot leave an intranet matters more than a few benchmark points.
Version History
- Xiaomi-OCR-0 listed: 0.8B document model open-sourced (2026-10-31): Form at listing: built on Qwen3.5-0.8B-Base and trained on about 170 million OCR and document samples, covering document parsing, key information extraction and OCR visual question answering in one model. Scores 96.83 on OmniDocBench v1.6, first among models under 1B parameters and 0.04 points from the top; 95.24 on Real5-OmniDocBench, 97.84 on UniMER-Test, 92.27 on PubTabNet and 83.2 averaged across five OCR-VQA benchmarks. Apache 2.0 weights and inference code, plus a browser demo, MCP service and Agent Skill, with a community Ollama build around 1.8GB. The technique already runs in Xiaomi's phone assistant for screen recognition and expense capture.