Xiaomi-OCR-0 Technical Report

Papers with Code Trending Papers
ocr document-parsing vision-language-model reinforcement-learning data-synthesis open-source

Summary

小米发布 Xiaomi-OCR-0 技术报告：一个基于 Qwen3.5-0.8B 的紧凑型视觉-语言模型，通过自动化数据引擎构建约 1.7 亿样本语料，并结合文本锚定与混合任务强化学习，在多个文档解析与 OCR 基准上取得 SOTA 成绩。

Compact OCR-specific vision-language models achieve strong document parsing performance, but often rely on costly supervision and focus primarily on visual-text reconstruction. We introduce Xiaomi-OCR-0, a unified 0.8B model for document parsing and OCR-centric understanding. We build an approximately 170M-sample OCR-centric corpus using an automated data engine that combines expert consensus, render-based verification, and targeted synthesis. Starting from Qwen3.5-0.8B, our progressive training recipe combines Q-Mask-based text anchoring, continued pretraining, and mixed-task reinforcement learning (Mix-RL). Xiaomi-OCR-0 achieves 95.24 on Real5-OmniDocBench, 96.83 on OmniDocBench v1.6, and 87.94 on Wild-OmniDocBench, while reaching an average score of 83.2 across five OCR-oriented VQA benchmarks. Ablations further show that, with sufficient parsing training, OCR-centric understanding supervision provides additional gains for document parsing. Homepage: https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0.
Original Article
View Cached Full Text

Cached at: 10/31/26, 03:11 AM

Paper page - Xiaomi-OCR-0 Technical Report

Source: https://huggingface.co/papers/2609.36136

Abstract

Xiaomi-OCR-0 is a compact vision-language model that unifies document parsing and OCR understanding through automated data synthesis, progressive training with text anchoring and reinforcement learning, and achieves state-of-the-art results on multiple benchmarks.

Compact OCR-specificvision-language modelsachieve strongdocument parsingperformance, but often rely on costly supervision and focus primarily on visual-text reconstruction. We introduce Xiaomi-OCR-0, a unified 0.8B model fordocument parsingandOCR-centric understanding. We build an approximately 170M-sample OCR-centric corpus using an automated data engine that combines expert consensus, render-based verification, and targeted synthesis. Starting from Qwen3.5-0.8B, our progressive training recipe combinesQ-Mask-based text anchoring,continued pretraining, andmixed-task reinforcement learning(Mix-RL). Xiaomi-OCR-0 achieves 95.24 onReal5-OmniDocBench, 96.83 onOmniDocBenchv1.6, and 87.94 onWild-OmniDocBench, while reaching an average score of 83.2 across fiveOCR-oriented VQAbenchmarks. Ablations further show that, with sufficient parsing training,OCR-centric understandingsupervision provides additional gains fordocument parsing. Homepage: https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0.

View arXiv pageView PDFProject pageGitHub228Add to collection

Get this paper in your agent:

hf papers read 2609\.36136

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### xiaomi-research/Xiaomi-OCR-0 Image-Text-to-Text• 0.9B• Updated2 days ago • 15.5k • 118

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.36136 in a dataset README.md to link it from this page.

Spaces citing this paper1

Collections including this paper2

Similar Articles

xiaomi-research/Xiaomi-OCR-0

Hugging Face Models Trending

Xiaomi-OCR-0 is a unified 0.8B vision-language model for document parsing and OCR-centric understanding, built from Qwen3.5-0.8B-Base, achieving #1 on Real5-OmniDocBench and outperforming larger models like Qwen3.5-2B on OCR VQA benchmarks.

Xiaomi-Lens-0 Technical Report

Hugging Face Daily Papers

Xiaomi-Lens-0 is a 4B multimodal model built on Qwen3.5 that predicts masks from its internal visual representations to improve fine-grained region reasoning, gaining 5.6 points on average over Qwen3.5 across RefCOCO-family referring comprehension benchmarks.

@poolmoon6: OCR is often needed. I use the one below, deployed locally. In real-world tests it outperforms Baidu OCR. On July 15, 2026, Alibaba released OvisOCR2. This is an end-to-end document parsing model based on Qwen3.5-0.8B, with only 0.8B parameters. On OmniDocBench v1…

X AI KOLs Timeline

Alibaba released OvisOCR2, an end-to-end document parsing model with 0.8B parameters based on Qwen3.5-0.8B, topping OmniDocBench v1.6 with a score of 96.58 — the first time an end-to-end approach has surpassed all pipeline-based approaches — and it is released under the Apache 2.0 license for commercial use.

OvisOCR3 Technical Report

Hugging Face Daily Papers

OvisOCR3 is a compact 0.8B vision-language model that unifies document parsing, general OCR, key information extraction, figure captioning, and cross-page table reconstruction via multi-teacher on-policy distillation and mixed-task reinforcement learning, achieving state-of-the-art results on benchmarks like OmniDocBench v1.6, OCRBench v2, and PureDocBench.

@GoSailGlobal: Current OCR processes multi-page documents page by page. Every time you turn a page, memory is reset. Today, Baidu quietly open-sourced a model on GitHub and HuggingFace called Unlimited OCR, inspired by how humans copy books: - When copying a book, you don't reread hundreds of pages every time you write a word...

X AI KOLs Timeline

Baidu has open-sourced the Unlimited OCR model, which uses a Reference Sliding Window Attention (R-SWA) mechanism to parse documents up to 32K context in a single pass, eliminating the need for page-by-page inference.