Kronos is an open-source foundation-model family for financial candlestick data, built to forecast market sequences using a tokenizer-plus-Transformer approach.
Kronos is a family of decoder-only foundation models focused on financial candlestick, or K-line, sequences. The repository presents it as the first open-source foundation model in this area and says it is trained on data from more than 45 global exchanges. It includes multiple model sizes, a live demo, and pretrained checkpoints available through the Hugging Face Hub.
The project targets forecasting and analysis of financial market time series, where the data are noisy and difficult to model with general-purpose time-series foundation models. It is designed to turn raw OHLCV-style market data into a form that can support a unified model for different quantitative tasks.
At a high level, Kronos uses a two-stage design. First, a specialized tokenizer converts continuous multi-dimensional candlestick data into hierarchical discrete tokens. Then a large autoregressive Transformer is pretrained on those tokens, so it can model and generate future market-sequence behavior; the repository also provides predictor classes that handle preprocessing, normalization, forecasting, and inverse normalization for users.
Its popularity appears tied to several signals in the README and metadata: it is positioned as a first-of-its-kind open-source model for financial candlesticks, it has a live demo, pretrained models on Hugging Face, and newly released fine-tuning scripts. The repository also notes an arXiv paper and an AAAI 2026 acceptance, which likely increased attention. The star count is already high and it gained stars today, suggesting active interest.
The README does not name direct competing repositories. It does, however, contrast Kronos with general-purpose time-series foundation models, implying that those are the closest broader alternative category. Within the project itself, the available alternatives are the different Kronos model sizes: mini, small, base, and large.
AI-explained · grounded in each repo's README