Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Xiaomi Releases Free 800M-Parameter OCR Model That Beats Qwen3.5-2B

Xiaomi has released a free OCR model, a neural network that reads text from scans and document photos and turns it into editable content. Xiaomi OCR-0 has just 800 million parameters (that is the model's size), yet in tests it beat Qwen3.5-2B, which has 2 billion parameters, 2.5 times as many. Text comes out as Markdown, formulas as LaTeX, and required fields as JSON. For contracts and invoices with confidential data, it is worth testing on your own documents.

AI-processed from TG: @bezsmuzi (Русский ИТ бизнес); edited by Hamidun News
Xiaomi Releases Free 800M-Parameter OCR Model That Beats Qwen3.5-2B
Source: TG: @bezsmuzi (Русский ИТ бизнес). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @bezsmuzi (Русский ИТ бизнес) ↗
Editor-in-Chief
Key metric #1Open source code, with the model available on Hugging Face
Local deploymentdocuments never go to third-party services
Key metric #3Converts text, tables and formulas into editable formats

Picture an archive of a thousand scanned contracts and invoices: to a computer, they are just images, but you need the numbers, dates and amounts in a table. This is the job of OCR, optical character recognition. Xiaomi has released Xiaomi OCR-0, a free OCR model with 800 million parameters. Parameters are the internal "settings" of a neural network, and the fewer there are, the cheaper and faster the model runs. According to the company, in tests it beat Qwen3.5-2B, a model with 2 billion parameters, which makes it 2.5 times larger. The code is open, and you can run the model yourself without sending documents to the cloud.

What Xiaomi OCR-0 Can Do

The model recognizes text in images and scans, extracts data from documents, and answers questions about their content. According to Xiaomi, it works in three main modes.

The first is converting content into an editable form. Text, tables and formulas from an image become something you can work with further, rather than a set of disconnected lines.

The second is converting to markup. A document becomes Markdown (plain text with headings and lists), and mathematical formulas are written in LaTeX (the standard format for formulas in scientific texts). This is handy for developers: the text can go straight into a knowledge base, be indexed for RAG, or be published in documentation.

The third is extracting the needed information into JSON. Say you have hundreds of invoices from different suppliers. Instead of writing a set of regular expressions for each template, you ask the model for the invoice number, date and amount, and get a structured response you can load into your accounting system.

Why the 800 Million Parameter Size Matters

The main figure in the announcement is the size. For OCR, it matters: the more compact the model, the cheaper and faster it is to process large volumes of pages, and the easier it is to run on an ordinary server or a powerful laptop. If a model is 2.5 times smaller than its competitor but does not fall behind it in tests, the hardware bill can drop noticeably.

An important caveat: the original announcement contains no results table, does not say which document sets the tests used, and does not show how large the margin is. For now, this is Xiaomi's own claim, not an independent verification. There is also no information about the license, beyond the fact that the source code is open.

What This Changes in Practice

Local deployment removes the most uncomfortable question: where the documents end up. For lawyers, accounting departments, medical providers, banks, and anyone handling personal data, sending scans to an external API often runs into security policy. If the model runs inside your own perimeter, there is nothing to negotiate.

The second benefit is cost. Cloud OCR services usually charge per page. On large archives, the sums add up quickly, while your own small model turns variable expenses into fixed ones: you pay for hardware, not for each document.

What to do right now:

  • Take 50–100 of your own real documents, not demo examples, and run them through the model;
  • Separately test the hard cases: handwritten notes, stamps over text, tables with merged cells, formulas;
  • Compare the results with your current solution on accuracy, speed and cost;
  • If you need JSON, fix the field schema in advance and validate the output with a validator, because any language model occasionally gets the format wrong.

Do not expect miracles on every document: a compact model may trail large systems on poor scans, unusual fonts or rare languages. But as a candidate for a first processing stage, especially where data cannot leave the premises, Xiaomi OCR-0 looks like a sensible option. You can find it on Hugging Face, and you can test it on your own documents in one evening.

Story

Key Questions Answered

What Xiaomi OCR-0 Can Do?

The model recognizes text in images and scans, extracts data from documents, and answers questions about their content. According to Xiaomi, it works in three main modes. The first is converting content into an editable form. Text, tables and formulas from an image become something you can work with further, rather than a set of disconnected lines. The second is converting to markup.

Why the 800 Million Parameter Size Matters?

The main figure in the announcement is the size. For OCR, it matters: the more compact the model, the cheaper and faster it is to process large volumes of pages, and the easier it is to run on an ordinary server or a powerful laptop. If a model is 2.5 times smaller than its competitor but does not fall behind it in tests, the hardware bill can drop noticeably.

What This Changes in Practice?

Local deployment removes the most uncomfortable question: where the documents end up. For lawyers, accounting departments, medical providers, banks, and anyone handling personal data, sending scans to an external API often runs into security policy. If the model runs inside your own perimeter, there is nothing to negotiate. The second benefit is cost. Cloud OCR services usually charge per page.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…