AnswersPublished24 sources
What Xiaomi-OCR-0 Can Do—and What Its Benchmarks Show
Announced in October 2026, Xiaomi OCR 0 is a 0.8B model for document parsing, key information extraction and visual question answering. It starts from Qwen3.5 0.8B and was trained on about 170 million OCR focused samples using text anchoring, continued pretraining and mixed task reinforcement learning.[1]
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Announced in October 2026, Xiaomi OCR 0 is a 0.8B model for document parsing, key information extraction and visual question answering.
- It starts from Qwen3.5 0.8B and was trained on about 170 million OCR focused samples using text anchoring, continued pretraining and mixed task reinforcement learning.[1]
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Xiaomi announce on October 30, 2026, about its open-source 0.8-billion-parameter Xiaomi-OCR-0 document understanding model, includi. Article summary: Xiaomi announced the release and open-sourcing of Xiaomi-OCR-0 on the evening of October 30, 2026: a lightweight, 0.8-billion-parameter model unifying document parsing, key-information extraction and visual question answ. Topic tags: general, academic, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake num
Xiaomi-OCR-0 is a compact vision-language model designed to do more than recognize text: it combines document parsing, key-information extraction and visual question answering. A news report published on October 31, 2026, says Xiaomi announced the model the previous evening. However, the project’s model card dates its technical report and weight release to September 30, so the October announcement does not appear to mark the first public availability of the model.39
8
One model for document parsing and OCR questions
Xiaomi describes the 0.8-billion-parameter model as a system for understanding document content and layout, including complex documents such as papers, tables and receipts. Its stated tasks include parsing documents, extracting key information and answering questions about visual content.39
1
The project’s developer materials also describe a browser demo and local tools: a GitHub repository includes an MCP service and an Agent Skill for OCR, PDF parsing, key-information extraction and visual question answering.30
31
32 These are developer interfaces around the model, rather than evidence that every workflow runs directly on a phone.
How Xiaomi says it was trained
The technical report describes Xiaomi-OCR-0 as starting from Qwen3.5-0.8B and training on an approximately 170-million-sample OCR-focused corpus. The data pipeline combines expert consensus, render-based verification and targeted synthesis.1
The model card outlines three training stages: Q-Mask text anchoring to align text with its position, continued OCR-focused pretraining, and mixed-task reinforcement learning using verifiable rewards on difficult examples.19 The available sources support this broad recipe, but do not establish every detailed data-generation method or reward definition mentioned in some descriptions.
Reported benchmark results
Xiaomi’s technical report and model-card materials give these headline results:1
3
| Benchmark or task | Reported score | Context |
|---|---|---|
| OmniDocBench v1.6 | 96.83 | Model-card materials rank it second in the cited comparison. |
| Real5-OmniDocBench | 95.24 | Xiaomi’s report and model-card materials give this score and claim first place. |
| Wild-OmniDocBench | 87.94 | Model-card materials rank it second in the cited comparison. |
| UniMER-Test | 97.84 | Formula recognition score reported in the model card. |
| PubTabNet | 92.27 | Table-recognition score reported in the model card. |
| Five OCR-focused visual-question-answering benchmarks | 83.2 average | The model card lists an average across DocVQA, InfoVQA, ChartQA, OCRBench and TextVQA. |
The Real5 score needs a caveat. The benchmark’s own leaderboard lists Xiaomi-OCR-0 at 92.73 overall, with scores of 93.73 for scanning, 92.76 for warping, 92.80 for screen photography, 92.68 for illumination and 91.69 for skew. That differs from Xiaomi’s reported 95.24; the available sources do not explain the discrepancy, so the two figures should not be treated as interchangeable.2
5
The five-benchmark VQA average is also a reported result, not a guarantee of performance on every document or question. The model-card materials compare the 83.2 average with Qwen3.5-2B at 80.9; comparisons with other models depend on the specific source and benchmark setup.3
14
Phone use and open developer tools
The October news report connects Xiaomi-OCR-0 with Super XiaoAi’s one-tap screen understanding. It gives examples such as finding a pickup code, reading a poster’s time and location, and extracting an amount and merchant from a bill.39
The project materials list model repositories on Hugging Face and GitHub, alongside a browser demo and local MCP and Agent Skill tools described in the GitHub project materials.25
9
30
31
32
What the evidence does—and doesn’t—establish
The sources support the main picture: Xiaomi-OCR-0 is a 0.8B document-understanding model, trained on a large OCR-focused corpus, with claimed capabilities spanning parsing, extraction and visual question answering. They also show that at least one headline benchmark figure is disputed across the model’s materials and the benchmark leaderboard.1
2
5
More specific claims—such as exact JSON output behavior for particular document types, detailed cross-region calculations, or every proposed phone workflow—are not established by the sources available here. The safest reading is to treat those as potential use cases, not verified capabilities.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Xiaomi-OCR-0 Can Do—and What Its Benchmarks Show"?
Announced in October 2026, Xiaomi OCR 0 is a 0.8B model for document parsing, key information extraction and visual question answering.
What are the key points to validate first?
Announced in October 2026, Xiaomi OCR 0 is a 0.8B model for document parsing, key information extraction and visual question answering. It starts from Qwen3.5 0.8B and was trained on about 170 million OCR focused samples using text anchoring, continued pretraining and mixed task reinforcement learning.[1]