What is Kronos?
Kronos is a family of decoder-only foundation models trained on financial candlesticks (K-lines) rather than text, by Yu Shi, Jian Li and five co-authors; the paper was accepted at AAAI 2026. A tokeniser turns open, high, low, close and volume bars into hierarchical discrete tokens, and an autoregressive Transformer is pre-trained on them; the arXiv paper puts the corpus at more than 12 billion K-line records from 45 exchanges.
Three sizes are open on Hugging Face under MIT: Kronos-mini (4.1M parameters, 2,048-token context), Kronos-small (24.7M) and Kronos-base (102.3M), both with a 512-bar context. Kronos-large (499.2M) is not released. The repo ships a KronosPredictor class for forecasts and a Qlib-based fine-tuning and backtest pipeline for the Chinese A-share market.
The GitHub repo has about 44,500 stars and an MIT licence. It has no tagged releases, and the last commit was in April 2026; Kronos-small is downloaded about 1.96 million times a month on Hugging Face.
Is Kronos worth it?
Kronos is worth having as a pre-trained starting point for market price bars, not as an oracle. General time-series foundation models tend to underperform on noisy financial data, the authors argue; Kronos instead learns a vocabulary of candlestick patterns from 45 exchanges and predicts the next bars the way a language model predicts words. The paper reports a 93% gain in RankIC over the leading general time-series model and 9% lower error on volatility forecasting. Those are the authors' own benchmark numbers.
The models are small by current standards. Kronos-small, the most downloaded, has 24.7M parameters, and predict_batch forecasts many assets in parallel on a GPU. Inputs are a pandas DataFrame with open, high, low and close (volume optional) and the timestamps you want forecast, and sampling settings let you draw several paths for a probabilistic view rather than one line.
The repo is candid about the gap to real use. Its fine-tuning example runs on Qlib's A-share data with a simple top-K backtest, and the README calls it a demonstration that ignores transaction costs, slippage, risk-factor neutralisation and portfolio construction. It also warns that many comments in the fine-tuning code were written by Gemini 2.5 Pro and may be wrong. The largest model, at 499.2M parameters, has not been released.
Maintenance is the other caution. There are no versioned releases, about 290 open issues, and no commits since April 2026, so it is a paper artefact that people use rather than a library that evolves. As a feature generator or baseline inside a quant team's own research stack it is worth an afternoon; anyone expecting it to beat the market from a notebook will be disappointed, and the authors say as much.
Where it earns its place
A quant researcher at a small systematic fund
Add a pre-trained forecast signal to an existing factor model
Run Kronos-base over daily bars for the universe with predict_batch, feed the forecast returns into the portfolio optimiser as one more signal, and judge it on a backtest that models costs. Free; one GPU covers inference.
A data scientist at a crypto exchange
Forecast short-horizon volatility for risk dashboards
Fine-tune the tokeniser and Kronos-small on the exchange's own 5-minute bars with the torchrun scripts, then sample several paths per pair to estimate a volatility band rather than a point forecast.
A finance PhD student
Compare a domain foundation model against classical baselines
Load the three open checkpoints from Hugging Face, reproduce the paper's zero-shot setting on a held-out market, and report where it fails, since the paper's gains are the authors' own measurements.
Before you commit
- Setup
- Clone the repo and pip install -r requirements.txt on Python 3.10+; load NeoQuasar/Kronos-small and its tokeniser from Hugging Face and call KronosPredictor.predict. Fine-tuning also needs pyqlib and locally prepared Qlib data.
- Integrations
- Pandas DataFrames in and out; weights on the Hugging Face Hub; Qlib for the data pipeline and backtest; optional Comet.ml logging in fine-tuning.
- Data and privacy
- Runs locally with no hosted calls beyond downloading weights. Code and the mini, small and base weights are MIT, so commercial use is allowed. No security advisories published.
- Cost at scale
- No paid edition. Inference on the 24.7M and 102.3M models needs modest hardware, a single GPU for batch work; fine-tuning scripts assume multi-GPU torchrun.
- Limits
- About 290 open issues; no commits since April 2026; 512-bar context on small and base; the backtest is a simplified demonstration by the authors' own description; Kronos-large is not available.
What’s good
- Pre-trained on 12 billion-plus K-line records from 45 exchanges
- Small open models, 4.1M to 102.3M parameters
- Probabilistic forecasts by sampling several paths
- MIT licence on code and released weights
What’s not
- No commits since April 2026 and no tagged releases
- Largest 499.2M model withheld
- 512-bar context on small and base models
- Backtest example ignores costs and slippage; benchmarks are self-reported
Best for
Quant researchers who want a pre-trained candlestick model as a baseline or feature source and will do their own backtesting properly.
Kronos alternatives
Kronos: common questions
- Is Kronos worth it?
- A credible research model that gives quants a pre-trained baseline for candlestick forecasting, not a trading system and not a maintained product. We rate it 3.8 out of 5.
- How much does Kronos cost?
- Free and open source; code and released weights are MIT.
- What are the downsides of Kronos?
- No commits since April 2026 and no tagged releases; largest 499.2M model withheld; 512-bar context on small and base models; backtest example ignores costs and slippage; benchmarks are self-reported.
- Who is Kronos best for?
- Quant researchers who want a pre-trained candlestick model as a baseline or feature source and will do their own backtesting properly.
- How much does Kronos cost for a team or in production?
- No paid edition. Inference on the 24.7M and 102.3M models needs modest hardware, a single GPU for batch work; fine-tuning scripts assume multi-GPU torchrun.
- Is Kronos safe for company data?
- Runs locally with no hosted calls beyond downloading weights. Code and the mini, small and base weights are MIT, so commercial use is allowed. No security advisories published.
- What are the best Kronos alternatives?
- The closest alternatives we have reviewed are Hugging Face Transformers.
Answers checked Oct 31, 2026.