Kronos: Tsinghua's open financial K-line foundation model, AAAI 2026 accepted

Language models read text; Kronos reads candlesticks. The Tsinghua-team open-source Kronos (MIT, 46,904 stars as of November 5, 2026) calls itself the first open-source foundation model for the “language” of financial markets: pre-trained on data from over 45 global exchanges and built for the uniquely high-noise character of financial K-lines. A Chinese write-up covered it at 39,000 stars; within a week it climbed to 46,904 — and on the academic side it was just accepted to AAAI 2026 (announced November 10), with the paper on arXiv (2508.02739).

Core features

• A two-stage framework: a specialized tokenizer first quantizes continuous, multi-dimensional OHLCV data into hierarchical discrete tokens; a large autoregressive Transformer is then pre-trained on those tokens — treating K-lines as a language to model.
• A four-tier model zoo: Kronos-mini (4.1M params, 2,048 context), small (24.7M) and base (102.3M) are open on Hugging Face; large (499.2M) is not.
• Live forecasting demo: the project site visualizes a BTC/USDT forecast for the next 24 hours.
• Fine-tuning scripts released (since August 17, 2025) to adapt Kronos to your own tasks and markets.
• Academic credentials: accepted at AAAI 2026; arXiv paper available.

Typical use cases

• A baseline for quant research: start from the open weights, fine-tune on your own data and compare.
• An experiment platform for market understanding: the K-line tokenization approach transfers to other high-noise time series.
• Teaching and demos: the live demo shows real forecast output, classroom-ready.

Quick start

Python 3.10+ and pip install -r requirements.txt; pull the model and tokenizer tier you need from Hugging Face. The BTC/USDT 24-hour forecast is viewable at shiyu-coder.github.io/Kronos-demo. A typical workflow: feed a historical K-line DataFrame, get OHLCV forecasts for the next bars, and use the August fine-tuning scripts to adapt to your own market or bar size. Training coverage spans crypto and US equities across the 45 exchanges; cross-market generalization is the core claim versus single-instrument models, with ablations in the paper.

Summary

Kronos fits quant researchers and financial-data engineers; anyone treating its output as a live trading signal should be cautious — the paper and repo provide forecasting capability, not a profit guarantee, and the high-noise nature of K-line prediction has not changed. Facts and limits: MIT, Python 3.10+; the large tier is closed; market-structure shifts after the training cutoff require fine-tuning. It marks the foundation-model playbook expanding into vertical data languages: build a tokenizer for K-lines first, then let a Transformer learn that language — now with an academic stamp of approval.