Kronos-base

Kronos-base

A 102M-parameter model for financial candlestick (K-line) data that forecasts price series and volatility and can generate synthetic data. Part of the Kronos family from NeoQuasar, trained on records from over 45 exchanges.

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-base is a 102M-parameter model from NeoQuasar for financial candlestick data, also called K-lines. The card describes the Kronos family as the first open-source foundation model for this kind of data, trained on records from over 45 global exchanges. It works in two stages. A tokenizer first converts the continuous OHLCV values of each candlestick into hierarchical discrete tokens. A decoder-only Transformer then predicts those tokens. The card says it forecasts price series and volatility and can generate synthetic data, with no extra training.

Who it suits: It suits quantitative researchers, data scientists, students of markets and developers of analytics tools.

Industry | Who uses it | How they use it | Example
Finance research | Quantitative researchers | Forecast price series from K-line data | Predict the next stretch of a candlestick series with KronosPredictor
Fintech | Developers of analytics tools | Add forecasts to dashboards | Show a forecast for a trading pair beside its history
Education | Students of markets and machine learning | Study a foundation model for market data | Compare the Kronos sizes on the same series

If you want trading advice or guaranteed results, you can skip it. A forecast is a model output, not a promise.

Where to use it:

• Python 3.10 or newer projects
• Notebooks and scripts that hold candlestick data with open, high, low, close and volume values
• Computers with room for about 409 MB of F32 weights
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Price forecasting | Quantitative researchers | A forecast of the next candlesticks
Volatility forecasting | Risk teams | An estimate of how far prices may swing
Synthetic data | Researchers | Generated candlestick series for testing

How to set it up: What you need first:

• Python 3.10 or newer
• The dependencies from the project's requirements.txt file on GitHub
• Candlestick data in open, high, low, close and volume form
• The Kronos-Tokenizer-base tokenizer that the model family table lists for this model

1. Install Python 3.10 or newer.
2. Install the dependencies from the requirements file.
3. Download the model and its tokenizer.
4. Use KronosPredictor to make a forecast.

pip install -r requirements.txt

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

How to use it: KronosPredictor prepares and normalizes your data, runs the prediction, then reverses the normalization. That takes you from raw data to a forecast in a few lines. For Kronos-base the max_context is 512, the longest sequence the model can read, so keep your lookback within that limit. The predictor trims longer input for you. The card walks through loading the tokenizer and the model and making a forecast, so follow those steps for the code.

Part | What it does
KronosPredictor | Prepares data, predicts, and reverses the normalization
Tokenizer | Turns candlestick rows into hierarchical discrete tokens
Kronos-base | The decoder-only Transformer that predicts the tokens
max_context | 512, the longest input the model can read

Example: You load a series of candlesticks for a trading pair, keep the lookback within 512 steps, and ask for a forecast. The card's live demo shows a forecast for the BTC/USDT pair over the next 24 hours.

For newcomers and experts: A good first try is a short series you already know, so you can judge the forecast against what the market did. Treat the result as an experiment.

For experts, the model learns by predicting the next token over a large corpus that spans many markets, with more than 12 billion K-line entries taken from 45 exchanges around the world. The tokenizer keeps both price dynamics and trade activity patterns. The card says the design targets the high-noise nature of financial data and works zero-shot across forecasting and generation tasks. The family has four sizes, and the largest is not yet public.

Candlestick data enters as OHLCV rows
The tokenizer turns it into discrete tokens
The Transformer predicts the next tokens
The predictor reverses the scaling to give a forecast

Good to know:

Title: The Kronos family
Model | Context length | Parameters
Kronos-mini | 2048 | 4.1M
Kronos-small | 512 | 24.7M
Kronos-base | 512 | 102.3M
Kronos-large | 512 | 499.2M, not yet public

• Context: Kronos-base reads up to 512 steps at a time.
• Noise: the card says the design targets the high noise of financial data, and forecasts can still be wrong.
• Demo: the card points to a live demo with a BTC/USDT forecast.
• Not advice: a forecast is not financial advice. Test it on your own data before you rely on it.
• License: MIT
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads files from Hugging Face. When you run it yourself, your price data stays on your machine.
• Not for: trading decisions based on forecasts alone.

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Parameters: 102M
• Weights: about 409 MB in F32; about 51 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 818,212
• Likes: 324
• License: mit
• Fine-tunes: 4
• Tags: safetensors, finance, candlestick, k-line, time-series-forecasting

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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