Kronos-mini

Kronos-mini

Forecasts financial candlestick series with a 4.1M-parameter foundation model trained on data from over 45 exchanges. For quant researchers and developers who want a small model for price forecasting experiments.

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-mini is the smallest public model in Kronos, a family of decoder-only foundation models for financial candlesticks, also called K-lines. A tokenizer first turns open, high, low, close and volume data into discrete tokens, and a Transformer learns from those tokens. The authors trained on over 12 billion records from more than 45 global exchanges. Kronos-mini has 4.1M parameters and a context length of 2048.

Who it suits: It suits people who want to test forecasting on market candlesticks with a small model that runs on modest hardware.

Industry | Who uses it | How they use it | Example
Finance | Quantitative researchers | Test price and volatility forecasts on candlestick data | Feed a history of daily candles and read the forecast
Fintech | Developers | Add a forecasting step to a dashboard or tool | Show a forecast chart next to live prices
Research | Machine learning researchers | Study a foundation model built for market data | Compare its zero-shot output with a baseline

If you want trading advice or promised returns, you can skip it, because the card makes no such claim.

Where to use it: You run it from Python, on Windows, macOS or Linux, through the KronosPredictor class in the project's code. The card also links a live demo of a forecast.

• Python 3.10 or newer
• The project's GitHub code, with its requirements file
• Windows, macOS or Linux with the huggingface_hub package

Use case | Who | What they get back
Price series forecasting | Quant researchers | Forecast candlesticks from past ones
Volatility forecasting | Analysts | A forecast of how much prices may swing
Synthetic data generation | Researchers | Artificial candlestick series for testing

How to set it up: You need Python 3.10 or newer and the dependencies from the project's GitHub repository. The weights are tiny, about 16 MB in F32.

• Python 3.10 or newer
• The dependencies listed in the project's requirements file
• About 16 MB of space for the weights

1. Install Python 3.10 or newer.
2. Install the dependencies from the requirements file.
3. Download the model files.

pip install -r requirements.txt
pip install -U huggingface_hub
hf download NeoQuasar/Kronos-mini

How to use it: Load the tokenizer and the model, wrap them in KronosPredictor, and pass it your candlestick history. The predictor handles preprocessing, normalization, prediction and the reverse normalization, so a forecast takes a few lines of code. The card's longer example is not shown in the part available here.

Tool | What it does
KronosPredictor | Runs preprocessing, normalization, prediction and inverse normalization
Kronos-Tokenizer-2k | The tokenizer that pairs with Kronos-mini
lookback | The amount of past data you pass in, ideally within the context limit

Example: You pass a history of BTC/USDT candlesticks to the predictor. It returns forecast candlesticks for the period after them. The card's live demo shows this kind of forecast for the next 24 hours.

For newcomers and experts: If you are new to this, think of a candlestick as one row of market data for a time period, and the model as a tool that continues the series. A first small step is to install the dependencies and run the card's forecast example on a small sample.

If you know the field, Kronos quantizes multi-dimensional K-line data into hierarchical discrete tokens, then pre-trains an autoregressive Transformer on them. It works zero-shot on forecasting tasks. Kronos-mini pairs with the 2k tokenizer, while small and base use the base tokenizer with a 512 context.

OHLCV candlestick history
Tokenizer makes hierarchical tokens
Transformer predicts the next tokens
Predictor reverses the normalization
Forecast candlesticks come out

Good to know: The family has four sizes, and the card says the largest one is not yet public. For the small and base models the maximum context is 512, and the predictor trims longer input automatically.

Title: Parameters by public Kronos model, in millions
Kronos-mini: 4.1 million
Kronos-small: 24.7 million
Kronos-base: 102.3 million

• License: MIT
• Cost to run: no API key or paid service is named, and the weights are about 16 MB in F32
• Data it reaches: the download comes from Hugging Face, and the model works on the market data you give it
• Not for: trading promises, since the card makes none, or the unreleased Kronos-large

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Parameters: 4.1M
• Weights: about 16 MB in F32; about 2 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 519,114
• Likes: 42
• License: mit
• Fine-tunes: 2
• Tags: safetensors, finance, candlestick, k-line, time-series-forecasting

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NeoQuasar/Kronos-mini

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