Kronos-Tokenizer-base

Kronos-Tokenizer-base

Converts candlestick price data into discrete tokens that the Kronos forecasting models read. It suits people who forecast financial markets in Python.

Made by NeoQuasar 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: Kronos-Tokenizer-base turns candlestick market data (open, high, low, close and volume) into hierarchical discrete tokens. The Kronos forecasting models read those tokens. It is the first stage of a two-stage design, and a Transformer is trained on the tokens in the second stage.

Who it suits: It suits people who forecast financial price series and want a ready-made encoder for candlestick data. Quant researchers and tool builders are the clearest fit.

Industry | Who uses it | How they use it | Example
Finance | Quantitative researchers | Encode candlestick history before forecasting with a Kronos model | Preparing a price series for a forecast run
Crypto | Developers of trading dashboards | Pair the tokenizer with a Kronos model for price forecasts | A BTC/USDT forecast like the live demo on the card
Academia | Students of time series models | Study how continuous prices become discrete tokens | Comparing this 4M tokenizer with the larger models that use it

Anyone who works with text, images or data that is not financial can skip it.

Where to use it:

• Python 3.10 or newer on Windows, macOS or Linux
• A PyTorch setup on your own computer
• The Kronos code on GitHub, which holds the KronosPredictor class
• The Hugging Face Hub, where the tokenizer files are hosted

Use case | Who | What they get back
Price forecasting | Quant researchers | Forecasts of future candlesticks from a Kronos model
Volatility forecasting | Risk analysts | Forecasts of how much a price may swing
Synthetic data generation | Machine learning engineers | Artificial K-line sequences for testing and research

How to set it up:

• Python 3.10 or newer
• The pip package manager
• Enough disk space for a tokenizer of about 16 MB

1. Install Python 3.10 or newer.
2. Install the Hugging Face hub package.
3. Install the dependencies from the requirements file in the Kronos GitHub repository.

pip install -U huggingface_hub

pip install -r requirements.txt

How to use it: Download the tokenizer files with one command. Then forecast with the KronosPredictor class, which handles preprocessing, normalization, prediction and inverse normalization for you. Keep your input within 512 steps, because that is the context length of Kronos-small and Kronos-base.

hf download NeoQuasar/Kronos-Tokenizer-base

Tool | What it does
hf download | Fetches the tokenizer files to your computer
KronosPredictor | Prepares your data, runs the forecast and converts the result back

Example: You hold a recent run of candlestick rows for a trading pair and want to know what comes next. You load this tokenizer with the Kronos-small model, hand the rows to KronosPredictor and ask for a forecast. The card's own demo does this for BTC/USDT over the next 24 hours.

For newcomers and experts: A tokenizer cuts data into small pieces a model can read. Here the data is candlesticks instead of words. A first small step is to download the files with the command above and then read the card's guide to making a forecast.

Underneath, continuous multi-dimensional price data is quantized into hierarchical discrete tokens, and an autoregressive Transformer is pre-trained on them. The card says this keeps both price movement and trading activity patterns. The tokenizer has 4M parameters. Kronos-mini uses a different one, Kronos-Tokenizer-2k.

Candlestick data (OHLCV)
Tokenizer makes discrete tokens
Kronos Transformer reads the tokens
KronosPredictor returns the forecast

Good to know:

• License: MIT.
• Cost to run: none named. The steps need no API key or paid service.
• Data it reaches: it fetches model files from the Hugging Face Hub. Your own candlestick data stays in your Python session.
• Not for: the largest Kronos model, with 499.2M parameters, is not yet public, so this tokenizer cannot run it.

Title: Size of each Kronos model in millions of parameters
Kronos-mini: 4.1 million
Kronos-small: 24.7 million
Kronos-base: 102.3 million

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Library: pytorch
• Parameters: 4M
• Weights: about 16 MB in F32; about 2 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,130,952
• Likes: 121
• License: mit
• Fine-tunes: 3
• Tags: pytorch, safetensors, finance, candlestick, k-line, time-series-forecasting

Links

• Model: https://huggingface.co/NeoQuasar/Kronos-Tokenizer-base
• Files: https://huggingface.co/NeoQuasar/Kronos-Tokenizer-base/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NeoQuasar/Kronos-Tokenizer-base

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