Xiaomi-OCR-0: a 0.8B open document model that parses pages into Markdown, tables, and JSON

Xiaomi's SeerRay team unveiled Xiaomi-OCR-0 on Oct 30: a unified 0.8B vision-language model built on Qwen3.5-0.8B-Base that turns document pages into Markdown, tables in OTSL, formulas in LaTeX, and key fields in JSON, plus OCR-centric question answering. Trained on ~170M OCR samples with a progressive recipe (Q-Mask anchoring, continued pretraining, mixed-task RL), it ships Apache-2.0 with weights, code, an MCP service, and an Agent Skill.

Why it matters

What you could build with it

Does it hold up?

Genuinely usable now: a Hugging Face demo, Ollama GGUF builds, and CPU support via Transformers exist, though the headline benchmark score is disputed against the public leaderboard, so real-world accuracy needs independent checks on messy documents.

Built with Xiaomi-OCR-0

• Xiaomi unveils Xiaomi-OCR-0, a 0.8B open source OCR model for documentsarticle · Independent writeup noting the full Oct 30 package (weights, demo, MCP service, Agent Skill) and flagging a benchmark discrepancy: 95.24 claimed vs 92.73 on the public Real5-OmniDocBench leaderboard.
• Xiaomi-OCR-0 GGUF quantizations for local inferencehuggingface · Community GGUF builds (down to 0.54 GB) with Ollama and llama.cpp quick-start, showing the model already runs on consumer hardware.
• Xiaomi Unveils Xiaomi-OCR-0, a 0.8B Open-Source AI Model for Reading Complex Documentsarticle · Covers the model's three tasks — parsing, key-information extraction, and visual question answering — and its Apache-2.0 release on GitHub and Hugging Face.

Learn more

technical deep dive →

First spotted on github: source.

More AI models