Kronos Is a 4.1-Million-Parameter Model Trained on Candlesticks. The Small Size Is the Argument.

In a week when a 753 billion parameter science model and a 397 billion parameter desktop agent both shipped, the repository climbing GitHub's trending list fastest in the finance category has 4.1 million parameters.
Kronos, by shiyu-coder, calls itself the first open-source foundation model for financial candlesticks. It was trained on data from more than 45 global exchanges, it is MIT licensed, and it was accepted at AAAI 2026. The repository has 47,000 stars and picked up roughly 3,000 this week, which is odd for a project whose last commit was in April.
The stars are worth thinking about, because they say something about what people actually want from AI right now that the frontier releases do not supply.
What it does
The premise is that a K-line chart is a language, and languages can be tokenized.
Kronos works in two stages. A tokenizer takes the continuous, multi-dimensional candle data, open, high, low, close, and optionally volume and amount, and quantizes it into hierarchical discrete tokens. A decoder-only Transformer is then pretrained autoregressively on those tokens, exactly the way a language model is pretrained on words.

That design choice is the whole bet. A general time-series foundation model usually treats a series as numbers and learns to regress the next value. Kronos treats a candle as a symbol and learns the distribution over the next symbol. The second framing lets you sample, which means you get a spread of futures instead of a point estimate, and it lets you borrow the entire toolkit that language modeling has built up over eight years.
The tokenizer is doing the hard work. A candle has five numbers that are not independent: high is always above open, close and low; low is always below them; volume correlates with range in ways that change by instrument. Quantizing that into discrete symbols without destroying the relationships is the part that took a paper.
The model zoo is unusually honest
Four sizes, three of them open.
Kronos-mini is 4.1 million parameters with a 2,048 token context. Kronos-small is 24.7 million and Kronos-base is 102.3 million, both with a 512 token context and a shared tokenizer. Kronos-large is 499.2 million and is not released.
Publishing the small models and withholding the large one is a defensible position and an increasingly common one. It gives researchers something to reproduce and gives the authors a commercial edge. It also means the best available open checkpoint is a 102 million parameter model, which is smaller than most phone keyboards.
The context lengths deserve attention because they bound what the model can see. Five hundred and twelve bars is not a lot. At five-minute intervals that is about 42 hours, under two trading days. At daily bars it is roughly two years, which is more reasonable. If your strategy depends on patterns that unfold over a quarter of hourly candles, this model cannot represent them in one pass.
How to run it
Python 3.10 or newer, then `pip install -r requirements.txt`. The weights come from Hugging Face under the NeoQuasar organization.
The interface is four lines:
```python tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base") model = Kronos.from_pretrained("NeoQuasar/Kronos-small") predictor = KronosPredictor(model, tokenizer, max_context=512) pred_df = predictor.predict(df=x_df, x_timestamp=x_timestamp, y_timestamp=y_timestamp, pred_len=120) ```
Your DataFrame needs open, high, low and close columns, with volume and amount optional. You pass a Series of historical timestamps and a Series of future timestamps, and the predictor handles normalization, forecasting and inverse normalization. It truncates inputs longer than the context window automatically, which is convenient and also a quiet way to lose information, so check your lookback.
There are fine-tuning scripts in the repo, one for the standard path and one that reads from CSV, plus a web UI and a set of examples. That is more complete than most research releases, which is part of why it keeps attracting stars.
What hierarchical tokens buy you
The word hierarchical in the tokenizer description is doing real work, and it is the part most people skip.
A flat tokenizer would quantize a candle into one symbol from a codebook of a few thousand. That compresses the five numbers into one index, and the relationships between them get smeared. A hierarchical tokenizer splits the job: a coarse token captures the broad shape of the candle and its position relative to recent range, and finer tokens capture the residual. Roughly, it is the difference between describing a chord as A minor and describing it as A minor with a sharp fourth in the bass.
Why it matters downstream is generation quality. When the model samples a coarse token, it has already committed to a region of candle space. The fine tokens then only need to decide where inside that region. Sampled sequences stay plausible for far longer, and plausible far horizons are the entire point of an autoregressive price model.
It also makes the tokenizer reusable. Kronos-small and Kronos-base share one tokenizer, so a team that fine-tunes the small model and later upgrades to the base does not have to rebuild its data pipeline.
The demo is a warning and a hint
The official live demo forecasts the BTC/USDT pair over the next 24 hours.
That is a strange choice for a model trained on 45 exchanges, and it is the right one for a demo. Crypto trades continuously, so nobody can accuse you of cherry-picking a market session. It is also the hardest possible test, because 24-hour crypto prices are close to a random walk and any model that looks good on them for a week will look bad the following week.
Which is the thing to keep in mind about this entire category. A foundation model for price series can be genuinely useful and still not tell you what to buy. The useful applications are downstream and specific: filling gaps in a data feed, generating plausible scenarios for risk simulation, detecting when a series has started behaving unlike its history, flagging a segment where the candle shapes are statistically improbable for that instrument. None of those require the model to predict direction.
The application that kills people is the obvious one. If you sample a thousand futures from Kronos and 60 percent of them drift up, that is a statement about the model's training distribution, not about tomorrow. Autoregressive models are very good at continuing the pattern they have seen, and financial markets have a habit of ending patterns precisely when enough capital is positioned for them to continue.
Why a small model is trending now
Three plausible reasons, and they are not exclusive.
The first is cost. A 102 million parameter model runs on a laptop. You can iterate on a feature pipeline in an afternoon without a GPU reservation, and the feedback loop for quantitative research is the same problem as the feedback loop for video generation: slow tools make cautious people.
The second is that the frontier got weird. Trillion-parameter models that need a data center to serve, agents that operate your desktop, world models that stream at 15 frames per second. A model that answers one narrow question on your own machine is a relief.
The third is the license. MIT, weights on Hugging Face, fine-tuning scripts included, no territory exclusions, no revenue threshold. For a quant team that wants to test whether candle tokenization works on its own instrument, that removes every reason not to try.
Whether the approach holds up is an empirical question that the repository does not settle. It publishes a model zoo and a demo, and the AAAI acceptance says the method passed peer review, which is not the same as saying it beats a well-tuned gradient-boosted model on your data. The honest test is a walk-forward evaluation on your own instrument against whatever baseline you already trust, with transaction costs included.
That test costs a weekend. The model costs nothing. That ratio is why the stars went up.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.