Xiaomi open-sources 0.8 billion parameter OCR model for document parsing

The new model provides developers with tools to extract structured data from complex documents like invoices and ID cards using native Markdown support.

ٹیکنالوجی31 اکتوبر، 2026تحریر Technology Desk

Xiaomi open-sources 0.8 billion parameter OCR model for document parsing

Xiaomi open-sourced its Xiaomi-OCR-0 model on 30 October, releasing the weights and inference code via GitHub and Hugging Face.

The 0.8 billion parameter model is designed to handle document parsing, key information extraction, and visual question answering. Xiaomi stated that the technology is built on the Qwen3.5-0.8B-Base model. It recognizes body text, tables, and math formulas to output structured results in reading order with native support for LaTeX and Markdown.

The system targets business forms, receipts, ID cards, and invoices by locating requested fields and returning them in JSON format. Users can perform reasoning over charts and documents, including numerical calculations and cross-region lookups. To ensure accuracy, the company filtered labels through a consensus of several expert OCR models that voted on text, table, and formula regions.

The development team employed a loop of predicting, re-rendering, and comparing to handle difficult formulas and tables. Xiaomi claimed the model scored 96.83 on OmniDocBench v1.6, ranking first among models under 1 billion parameters and second overall. Xiaomi noted that Xiaomi-OCR-0 outperformed HunyuanOCR-1.5 at 76.5 and 8B MiniCPM-V-4.5 at 82.6, though it scored lower than Qwen3.5-4B at 84.9.

The technology currently operates within the Super XiaoAi assistant on Xiaomi phones to pull meeting times, bill details, queue numbers, and pickup codes from screens. Developers can access a browser demo, an MCP service, and an Agent Skill toolkit provided by the company.

متعلقہ خبریں

ٹیکنالوجی2 گھنٹے پہلے

AgiBot proposes Worken unit to measure robot productivity in China: Peng Zhihui

The unit distinguishes uptime from value, aiming to reduce the need for individual engineers to manage each robot as operating costs fall.

ٹیکنالوجی4 گھنٹے پہلے

InSpatio system generates robot trajectories in hundreds of parallel environments

The technology addresses the lack of data variety in robot training by using high-speed simulations to create complex motion plans.

ٹیکنالوجی5 گھنٹے پہلے

Tencent Cloud launches integrated platform for physical AI applications

The new solution aims to streamline the physical AI engineering chain by consolidating fragmented tools into a single platform to reduce costs and time.

ٹیکنالوجی9 گھنٹے پہلے

Creative groups file amicus brief in New York Times lawsuit against OpenAI

The brief argues that existing industry infrastructure can clear rights at scale and that fair use claims would undermine recent labor agreements.

ٹیکنالوجی11 گھنٹے پہلے

OpenAI, Google, Microsoft and Anthropic to adopt AI scraping opt-out standard

The new policy token allows publishers to manage AI training preferences across multiple crawlers while maintaining search engine visibility.

ٹیکنالوجی14 گھنٹے پہلے

Binance secures $3bn of digital assets after $6.3bn cyber hack

The breach is linked to a North Korean hacking unit and involved a self-destruct mechanism that destroyed traces of the attack.