Kronos

Kronos is the first open-source foundation model for financial candlesticks (K-lines), with 42,000+ stars on GitHub, trained on data from more than 45 global exchanges. It treats K-line sequences as a language the model learns to forecast and represent, giving quantitative researchers an open, research-grade starting point instead of training a model from scratch.

Overview

Kronos was published by shiyu-coder with its paper on arXiv (2508.02739) and accepted for AAAI 2026, which puts it in rare company among open-source financial ML projects: peer-reviewed work with reproducible code. The repository ships the full model pipeline plus released fine-tuning scripts, so you can adapt Kronos to your own markets, timeframes, and tasks rather than being locked to the pre-trained setting. Weights are published on Hugging Face (NeoQuasar), and a live demo lets you probe the model before you clone anything.

Key Features

• First of its kind. A foundation model purpose-built for candlestick (K-line) data, not a repurposed language or vision model.
• Trained on 45+ global exchanges. Broad cross-market coverage is baked in from pre-training, reducing single-market overfitting.
• Peer-reviewed. Accepted at AAAI 2026 with the paper public on arXiv, so the methodology is open to scrutiny.
• Fine-tuning scripts included. Adapt Kronos to your own assets, horizons, and objectives with the released training code.
• Live demo and Hugging Face. Try the model in the browser first, then pull the weights and code.
• Fully open. Model, weights, code, and paper are all public — research-grade transparency end to end.

Pricing

Free. Kronos is open-source research software: the model, weights, code, paper, and demo are all available at no cost.

Who Is It For

Quantitative researchers, fintech engineers, and ML practitioners building forecasting, backtesting, or market-analysis pipelines — anyone who would otherwise spend weeks training a candlestick model from zero now has a peer-reviewed foundation to start from.