Typhoon OCR V2 0.8B

Typhoon OCR V2 0.8B

Parses Thai and English document pages into ordered layout elements with categories, boxes and text, using 0.8B parameters. For developers who digitize Thai paperwork.

Made by typhoon-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Typhoon OCR V2 0.8B is a document parser for Thai and English with 0.8B parameters. Send it a picture of one page and it lists every element it finds, ordered the way a person would read the page. Each entry carries a type label, a box that marks where it sits and the words inside. Table content arrives as HTML markup, math as LaTeX and the remaining text as Markdown.

Who it suits: Developers and teams who turn Thai or English page images into structured data and want a small model to do it.

Industry | Who uses it | How they use it | Example
Government | Records offices | Digitize Thai forms and letters | A scanned form becomes ordered text with boxes
Finance | Back-office teams | Pull tables out of statements | A statement table comes back as HTML
Research | Academics | Extract text and formulas from papers | A formula comes back as LaTeX
Software | Developers | Add page parsing to an app | An app reads a photographed page

You can skip it if you need to answer questions about a document. The card sends you to the larger 2B version for that.

Where to use it:

• Hugging Face Transformers, which the card says is unaffected by the vLLM issue
• vLLM, version 0.30.0 or newer
• Ollama, llama.cpp and MLX, through other builds the card lists
• The hosted Typhoon API, as the model typhoon-ocr-v2

Use case | Who | What they get back
Full-page parsing | Developers | Layout elements in reading order with categories and boxes
Table extraction | Analysts | Tables as HTML
Formula extraction | Researchers | Formulas as LaTeX

How to set it up:

• Python and pip on your computer
• vLLM 0.30.0 or newer, or Transformers, to run the model
• Optionally the typhoon-ocr2 Python package

1. Download the model files with the hf command.
2. Install vLLM 0.30.0 or newer. On an older vLLM, run in eager mode.
3. Follow the usage section of the model card to send it a page image.

pip install -U huggingface_hub
hf download typhoon-ai/typhoon-ocr2-0.8b

pip install typhoon-ocr2

How to use it: Send the model a page image and read back the structured elements. With an older vLLM, add the eager flag when you serve the model. The card also lists an Ollama build.

Tool | What it does
vllm serve with --enforce-eager | Runs the model on older vLLM versions at full accuracy
typhoon-ocr2 package | Python helper for calling the model
ollama run SuraponOCR/typhoon-ocr2-0.8b | Runs the model through Ollama

ollama run SuraponOCR/typhoon-ocr2-0.8b

Example: You send a photo of a Thai invoice page. The model returns a list of elements, such as a title, text blocks and a table. Each has a category and a box, and the table text is in HTML.

For newcomers and experts: If you are new, try it on one clean page image and read the output in order. Use it only for whole-page parsing, since the card says it is much weaker on other tasks.

If you know the field, it is a fine-tune of OvisOCR2 trained with supervised fine-tuning and then GRPO, with rewards that give zero to runaway outputs. On MDPBench the card reports 77.4 against 66.0 for Typhoon OCR 1.5. Some English and Chinese accuracy was lost against the base model.

Page image
Typhoon OCR V2 0.8B
Elements in reading order
Boxes, categories and text

Title: MDPBench score
Typhoon OCR 1.5: 66.0
Typhoon OCR V2 0.8B: 77.4

Good to know:

• License: Apache 2.0, and the card says use is also subject to the OpenTyphoon Terms and Conditions
• Cost to run: your own hardware, or the hosted Typhoon API, whose price the card does not give
• Data it reaches: the page images you give it and the files you download
• Not for: answering questions about documents, which the card assigns to the 2B model
• Careful: older vLLM versions in compiled mode lowered accuracy in the card's tests

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 853M
• Weights: about 1.7 GB in BF16; about 426 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 485
• Likes: 8
• License: apache-2.0
• Languages: th, en
• Based on: ATH-MaaS/OvisOCR2
• Quantized versions: 4
• Fine-tunes: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, document-parsing, layout-analysis, table-recognition, thai, conversational

Links

• Model: https://huggingface.co/typhoon-ai/typhoon-ocr2-0.8b
• Files: https://huggingface.co/typhoon-ai/typhoon-ocr2-0.8b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download typhoon-ai/typhoon-ocr2-0.8b

Then follow the usage in its model card. Nothing in this kit runs by itself.