Kronos: A Chinese Translation of the Financial K-Line Foundation Model

Kronos is a decoder-only foundation model for financial candlestick (K-line) sequences. The original repository, shiyu-coder/Kronos, has 47,035 stars and is written in Python. A Chinese translation of the project now exists at yangshun2005/Kronos-cn, released today. It has zero stars at the time of writing, and I won't pretend otherwise.

This article covers what Kronos does, how the Chinese version is structured, and where the translation actually helps. It is not a benchmark review — I have not run the model, and the numbers below come from the project's own documentation.

What Kronos is

The project describes itself as the first open-source foundation model designed specifically for financial K-lines, pretrained on data from more than 45 global exchanges. It was accepted at AAAI 2026, and the paper is on arXiv (2508.02739).

The architecture is a two-stage framework:

1. A dedicated tokenizer quantizes continuous multi-dimensional OHLCV data into hierarchical discrete tokens.
2. A large autoregressive Transformer is pretrained on those tokens.

The stated motivation is that general time-series foundation models don't handle the noise profile of financial data well. Kronos is built around that noise rather than treating it as an afterthought.

Released checkpoints come in several sizes — Kronos-mini, Kronos-small, and Kronos-base — available from the Hugging Face Hub. Parameter counts range from 4.1M to 499.2M. Kronos-large (499.2M) is marked as not open-sourced in the model table, so you can't download it.

Why a Chinese version matters here

The original README is thorough, but it is long, and a fair amount of the practical detail — context length limits, required DataFrame columns, the fact that you must supply your own timestamp Series — is scattered through it. For a Chinese-speaking developer evaluating the model, that's friction before any code runs.

The Chinese version (Kronos-cn) does two things:

• It translates the original README body, aiming to stay faithful to the source. The original English README is kept as README_en.md.
• It adds a "中文快速上手" (Chinese quick start) section at the top. The translator explicitly labels this as supplementary guidance, not part of the original document.

That second part is the useful bit. It's a condensed path to a running prediction, written for people who want to get the model loaded before reading the full documentation.

The translation date is listed as 2026-11-07. The license is MIT, same as upstream.

Getting it running

The quick start section gives a short path. Requirements are Python 3.10+ and the dependencies in requirements.txt:

pip install -r requirements.txt

Those dependencies include torch>=2.0.0, einops==0.8.1, huggingface_hub==0.33.1, pandas==2.2.2, numpy, matplotlib==3.9.3, tqdm==4.67.1, and safetensors==0.6.2.

Load the tokenizer and model from the Hub:

from model import Kronos, KronosTokenizer, KronosPredictor

tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
model = Kronos.from_pretrained("NeoQuasar/Kronos-small")

Instantiate the predictor. For Kronos-small and Kronos-base, max_context is 512:

predictor = KronosPredictor(model, tokenizer, max_context=512)

Then prepare input. df must be a pandas DataFrame with open, high, low, close columns; volume and amount are optional. You also need a historical timestamp Series and a future timestamp Series:

import pandas as pd

df = pd.read_csv("./data/XSHG_5min_600977.csv")
df['timestamps'] = pd.to_datetime(df['timestamps'])

lookback = 400
pred_len = 120

x_df = df.loc[:lookback-1, ['open', 'high', 'low', 'close', 'volume', 'amount']]
x_timestamp = df.loc[:lookback-1, 'timestamps']
y_timestamp = df.loc[lookback:lookback+pred_len-1, 'timestamps']

The final predict call is where the translated quick start stops — the excerpt I have is truncated at that point, so I can't show the exact arguments or return values. The full README body covers it.

The pitfalls the translation surfaces

The quick start section has a "常见坑" (common pitfalls) list. Three of these are worth repeating because they're easy to miss when skimming the original:

Context length. Kronos-small and Kronos-base have max_context of 512. Keep lookback at or below that. KronosPredictor truncates longer contexts automatically, which means you can silently lose history if you don't check.

Model download. The first run pulls the tokenizer and weights from Hugging Face. In mainland China this often needs a mirror or proxy configured in advance. The translation notes this; the original documentation doesn't emphasize it.

Input format. The DataFrame column requirements and the separate timestamp Series are both mandatory. This is a common source of errors for anyone used to libraries that infer timestamps from an index.

Also worth flagging: Kronos-large is not available. If you were planning around the largest checkpoint, plan around base instead.

What the translation doesn't cover

The README excerpt I have doesn't state operating system requirements, and it doesn't give specific hardware requirements beyond noting that model sizes range from 4.1M to 499.2M parameters so you can pick according to your compute. If you need concrete VRAM figures, you'll have to measure them yourself.

There's also an online demo that visualizes predictions — the page shows a BTC/USDT 24-hour forecast — if you want to see output before installing anything.

Fine-tuning scripts were released on 2025.08.17, so adapting Kronos to your own task is supported. The translation's quick start focuses on inference; for fine-tuning you'll want the full README.

Is the Chinese version worth using

If you read Chinese and want to evaluate Kronos quickly, the translated quick start saves you a pass through a long English document. The translation is honest about its own scope — it separates translated content from added guidance, and invites issues or PRs for inaccuracies. That's the right structure for this kind of fork.

If you read English comfortably, there's little reason to switch; the upstream repo is the canonical source and will stay current. The Chinese version is a snapshot dated 2026-11-07, and translations drift.

One caveat: this is a fresh translation with no community review yet. My read of the quick start is that it's accurate against the original, but my sample is one document, and terminology choices in a financial ML context can be contentious. Treat it as a convenience layer, not a replacement for the upstream README when something looks off.

The original project is at shiyu-coder/Kronos. The Chinese version is at yangshun2005/Kronos-cn.

Chinese documentation: https://github.com/yangshun2005/Kronos-cn

I'm an independent enterprise AI delivery consultant — private LLM deployment, RAG, agent workflows, AIOps. If your team is stuck somewhere in that stack, my inbox is open.