Shiyu Coder · 2026-01-15 · notable
Kronos — Foundation Model for Financial K-Line Sequences
Kronos is the first open-source foundation model specifically designed for financial candlestick (K-line) sequences. Trained on data from 45+ global exchanges, accepted at AAAI 2026. Models range from 4.1M to 499.2M parameters.
The first foundation model that treats financial candlestick data as a specialized language.
Key specs
| License | MIT |
|---|---|
| GitHub stars | 17.7k |
| Model sizes | 4.1M - 499.2M params |
| Exchanges trained on | 45+ |
What is it?
Kronos is an open-source foundation model family engineered specifically for financial K-line (candlestick) sequences. It treats OHLCV (open, high, low, close, volume) data as a specialized language using hierarchical tokenization. The model lineup ranges from Kronos-mini (4.1M params) to Kronos-large (499.2M params). Accepted at AAAI 2026.
How does it work?
A two-stage framework: a specialized tokenizer quantizes continuous OHLCV data into hierarchical discrete tokens, then an autoregressive Transformer is pre-trained on these sequences. This departure from general-purpose time-series models handles the unique high-noise characteristics of financial data. Supports context lengths from 512 to 2048 tokens, batch processing, and temperature/nucleus sampling.
Why does it matter?
Financial time-series forecasting has been dominated by ARIMA variants and general-purpose models. Kronos is the first domain-specific foundation model for trading data, potentially enabling transfer learning across markets just as language models transfer across domains. The AAAI acceptance validates the approach academically.
Who is it for?
Quantitative researchers, trading system developers, fintech companies.
Try it
pip install kronos-fin or download from Hugging Face.