Kronos: an open foundation model that reads candlesticks

Kronos is the first open-source foundation model built specifically for financial candlesticks, the OHLCV K-lines every trader stares at. It was trained on data from more than 45 global exchanges, and instead of being a chatbot bolted onto a market, it treats price history like a language: a tokenizer chops the multi-dimensional data into discrete tokens, and an autoregressive Transformer predicts what comes next.

The model family runs from a tiny 4.1M-parameter version up to about 500M, with context up to 2,048 tokens, so you can run it on a laptop or push it harder on a GPU. It's not a trading agent that clicks buy and sell, it's the forecasting brain you'd wire underneath one. Give an agent a Kronos head and it stops guessing at charts from raw text and starts reasoning over a representation actually trained on markets.

Why it keeps trending: everyone building finance agents has been missing this exact piece. There are great language models and great general time-series models, but almost nothing that's open, market-native, and small enough to actually deploy. Kronos fills that hole, and the fact that it's back on GitHub's daily trending months after release tells you people are still discovering they need it.

Repo and weights: https://github.com/shiyu-coder/Kronos