Kronos is an open-source foundation model that treats the price action of global financial markets as a language to be learned. Released under the MIT license by an independent academic group and accepted to AAAI 2026, the project has gathered 25,000+ GitHub stars and 4,300+ forks since its public release, making it the most-starred open financial foundation model on GitHub. Where general-purpose time-series models such as TimesFM or Chronos try to fit any temporal signal, Kronos is built end-to-end for the statistical quirks of K-line, or candlestick, data. ## Why a Financial Foundation Model Financial time series are notoriously hostile to off-the-shelf forecasting models. Returns are heavy-tailed, volatility clusters, regimes shift, and microstructure noise dominates short horizons. Kronos addresses this by pre-training a decoder-only transformer on candlestick sequences from 45+ global exchanges, then exposing a small family of open checkpoints that can be fine-tuned on a specific market, asset class, or trading horizon. The same model that produces a BTC/USDT forecast in the project's live demo can be adapted to forecast A-shares on the Shanghai exchange with a few hundred lines of code and the included Qlib integration. ## Architecture: Tokenizing OHLCV The innovation that sets Kronos apart from generic time-series transformers is its tokenizer. Raw OHLCV data, the open, high, low, close, and volume tuples that define a candle, are continuous and multi-dimensional. Kronos quantizes these into hierarchical discrete tokens that preserve the relative structure of each bar while exposing the sequence to standard autoregressive modeling. The downstream transformer is a plain decoder-only stack, which means it inherits the well-understood scaling, inference, and serving infrastructure of modern LLMs and can be quantized, batched, and served with the same tooling teams already use for language models. ## Open Model Family Kronos ships three open checkpoints. Kronos-mini at 4.1M parameters with a 2048-token context targets edge and embedded analytics. Kronos-small at 24.7M parameters and Kronos-base at 102.3M parameters both use a 512-token context and target serious quantitative research workflows. A fourth proprietary checkpoint, Kronos-large at 499.2M parameters, is held back from the open release. All open models are distributed on the Hugging Face Hub and integrate cleanly with the Hugging Face transformers library and Qlib, Microsoft's open-source quantitative platform. ## What Comes In The Box The repository is deliberately practical. A live BTC/USDT forecasting demo lets users see model behavior on real data before downloading anything. A batch-prediction pipeline parallelizes inference across many tickers at once, which matters for portfolio-scale workflows. A complete fine-tuning recipe ships with backtesting hooks, so a quant researcher can take a base checkpoint, fine-tune on their universe, and run a walk-forward backtest without leaving the repo. Documentation is provided in eight languages, reflecting an unusually international contributor base for an academic release. ## Use Cases Kronos is intended for forecasting tasks across timeframes from intraday to swing-trading horizons, quantitative strategy research, multi-asset portfolio modeling, and risk and regime-detection work. The decoder-only design also makes it a natural backbone for synthetic-market generation and scenario stress testing, since the model can autoregressively roll out plausible future candle sequences conditioned on the recent past. ## Limitations Kronos forecasts price action, not causal market structure, so it carries the usual caveats of any foundation model in finance. The open checkpoints are trained on historical data through a fixed cutoff and need fine-tuning to adapt to new regimes. The largest and most capable variant is not part of the open release, which limits what researchers can reproduce from the published benchmarks. As with any quantitative model, deploying Kronos directly to live trading without risk overlays, position sizing logic, and human review is a poor idea. The MIT license is permissive, but local regulations around algorithmic trading still apply.