Xiaomi-OCR-0
Xiaomi's 0.9B Apache 2.0 OCR model on Qwen3.5-0.8B for parsing, extraction and OCR questions; first on Real5-OmniDocBench by its own table.
In short. Xiaomi-OCR-0 is a very small AI model from Xiaomi that reads documents and also answers questions about the text in an image. At under one billion parameters it fits on almost any GPU and many laptops. In Xiaomi's own tests it matches or beats larger OCR models, and the license allows commercial use.
Summary
Xiaomi's research team released Xiaomi-OCR-0 with a technical report on 30 September 2026. It is a 0.87-billion-parameter vision-language model, about 1.75 GB in BF16, trained from Qwen3.5-0.8B-Base on roughly 170 million OCR-centered samples. One model covers document parsing, OCR-focused question answering and key information extraction. The recipe has three stages: Q-Mask text anchoring that aligns text with its position, OCR-centered continued pretraining, and mixed-task reinforcement learning with verifiable rewards.
The model card reports 96.83 on OmniDocBench v1.6, ranked second in its comparison, 95.24 on Real5-OmniDocBench, ranked first, and 87.94 on Wild-OmniDocBench, ranked second. In the same table PaddleOCR-VL-1.6 scores 96.33 on OmniDocBench v1.6 and GLM-OCR scores 95.22. It also reports 97.84 on the UniMER-Test formula benchmark, 92.27 on PubTabNet tables, and an average of 83.2 across five OCR-centered question-answering benchmarks, above Qwen3.5-2B at 80.9. All results are Xiaomi's own.
The training data engine combined annotations from a pool of existing OCR models, including PaddleOCR-VL-1.6, MinerU2.5-Pro, GLM-OCR, dots.mocr and HunyuanOCR-1.5, and kept the cases where several of them agreed. The weights are under Apache 2.0, and community GGUF conversions already exist for llama.cpp-style runtimes.
Good for
- Document parsing and OCR question answering on small GPUs and laptops
- Pipelines that need a single tiny model for parsing, extraction and questions
- Research on training small OCR models with consensus-labeled data
- Commercial use under Apache 2.0
Watch out for
- Rankings and scores come from Xiaomi's own comparison tables
- It is a first release, labeled version 0, and its adoption is still small
- It is trained on annotations produced by other OCR models, so check their terms if that matters to you
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Xiaomi-OCR-0 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Xiaomi-OCR-0 BF16 | 873M | 2.01 GB | 1 GB | 0.602 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source OCR and document models in 2026
- Eleven open OCR models in a year: small and hard to compare
Frequently asked questions
- What is Xiaomi-OCR-0?
- Xiaomi-OCR-0 is a very small AI model from Xiaomi that reads documents and also answers questions about the text in an image. At under one billion parameters it fits on almost any GPU and many laptops. In Xiaomi's own tests it matches or beats larger OCR models, and the license allows commercial use.
- Can I use Xiaomi-OCR-0 commercially?
- Yes. Xiaomi-OCR-0 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Xiaomi-OCR-0 open source?
- Xiaomi-OCR-0 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Xiaomi-OCR-0 on my own computer?
- Yes. The smallest variant, Xiaomi-OCR-0, needs about 0.602 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Xiaomi-OCR-0 released?
- Xiaomi-OCR-0 was released by Xiaomi on September 30, 2026.
- How long a context window does Xiaomi-OCR-0 have?
- Xiaomi-OCR-0 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Xiaomi-OCR-0 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone | |
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| MiniCPM-OCR 1.0 OpenBMB | 1.3B | 256K | Apache 2.0 (Permissive) | 0.897 GB Laptop or phone | |
| HunyuanOCR Tencent | 1.12B | 128K | Tencent Hunyuan (Conditional) | 0.773 GB Laptop or phone |