khammoon

khammoon

Splits Thai text into morphemes, the smallest pieces that still carry meaning, with a 4 MB model that runs on a CPU, for Thai language researchers and developers.

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: khammoon is a small Thai text splitter. It cuts Thai text into morphemes, the smallest pieces that still carry meaning, following the annotation guideline in the project's linguistics folder. Usual Thai word splitters keep compounds whole, while khammoon breaks them into their meaningful parts. The model works on characters, weighs about 4 MB with roughly 1.06M parameters, and runs on a CPU with no GPU and no large language model.

Who it suits: Thai language engineers and researchers who prepare text for tagging, entity recognition or sentiment work will find it fits. Linguists who follow the same morpheme guideline can use it too.

Industry | Who uses it | How they use it | Example
NLP research | Thai NLP researchers | Split text into morphemes before training a tagger | Preparing a corpus for part-of-speech tagging
Search and text tools | Software engineers | Break Thai strings into small units on a plain CPU | Splitting product names in a catalogue
Linguistics | Linguists | Apply one morpheme definition across a large corpus | Checking how a compound word breaks down
Education | Language-tool builders | Show learners the meaningful parts of a word | Splitting a compound into its parts

If you work with languages other than Thai, you can skip it, since the card lists only Thai.

Where to use it:

• Python scripts, through the khammoon package
• The command line, where the khammoon tool reads text line by line
• Programs that load ONNX models, since Hugging Face lists the ONNX format
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Prepare training text | NLP researchers | A list of morpheme tokens for each sentence
Process a corpus | Data engineers | Token lists for many texts in one batch
Quick check in a terminal | Linguists | The pieces of a phrase shown on one line

How to set it up: You need Python with pip. The card says the base package needs only numpy, and onnxruntime is an optional extra for speed.

• Python with pip, on Windows, macOS or Linux
• The khammoon package, which needs only numpy
• Optional: onnxruntime, which the card says is about 2.4 times faster with the same output
• About 4 MB of space for the model

1. Install the khammoon package with pip.
2. Optionally add onnxruntime for the faster backend.
3. Import it in Python and split a Thai string.

pip install khammoon
pip install -U huggingface_hub
hf download typhoon-ai/khammoon

How to use it: Call the segment function on a Thai string and it returns a list of pieces. For many texts, hand a whole list to a Segmenter in one go, which runs faster than separate calls. The command line tool reads standard input line by line.

khammoon "โรงเรียนสอนคนตาบอด"

Tool | What it does
segment | Splits one Thai string into a list of morphemes
Segmenter | Loads the bundled model for repeated or batch calls
threshold | A lower value gives more boundaries and better recall
backend | Picks onnx or numpy, and auto uses onnx when onnxruntime is installed
khammoon command | Splits text from the terminal or from piped input

Example: You run the command line tool on the phrase from the card. It prints six pieces: โรง, เรียน, สอน, คน, ตา and บอด.

For newcomers and experts: A morpheme is the smallest piece of a word that still has meaning. A good first step is to run the one-line command on a short Thai phrase and look at the pieces.

The model has about 1.06M parameters and learned from labels made by a large language model that segmented text in two passes. The card says khammoon makes about 5 times fewer errors than that model on its test set. It never alters the text, since joining the pieces always gives back the input, each space is its own piece, and splits fall only between Thai character clusters. On the card's test of 2,000 documents, boundary F1 is 0.992, against 0.925 for AttaCut. The card notes that older tools split into words, so their recall is low by design. These scores come from the authors.

You pass in a Thai string
The model scores each boundary between character clusters
Boundaries above the threshold become splits
You get back a list of morpheme tokens

Title: Boundary F1 on the card's test set, higher is better
PyThaiNLP newmm: 0.802 F1
deepcut: 0.921 F1
AttaCut: 0.925 F1
LLM segmenter: 0.962 F1
khammoon: 0.992 F1

Good to know:

• License: cc-by-nc-sa-4.0, as listed on Hugging Face, so check its terms before any paid use
• Cost to run: none named, and it needs no GPU
• Data it reaches: the Thai text you give it, and the source names no web or account access
• Speed on one CPU thread: about 30,000 characters a second with ONNX and about 13,000 with numpy
• Not for: languages other than Thai, since the card lists only Thai
• Older tools are scored here against the morpheme definition, not at their own word-splitting task

On Hugging Face, 13 November 2026

• Task: token classification
• Library: onnx
• Downloads in the last 30 days: 0
• Likes: 4
• License: cc-by-nc-sa-4.0
• Languages: th
• Tags: onnx, thai, word-segmentation, morpheme-segmentation, token-classification

Links

• Model: https://huggingface.co/typhoon-ai/khammoon
• Files: https://huggingface.co/typhoon-ai/khammoon/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download typhoon-ai/khammoon

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