Mhondoro-base

Mhondoro-base is part of the Mhondoro family of open-source foundation models for financial time-series forecasting, released by Louis Chifura.

Mhondoro models tokenize OHLCV candlestick (K-line) sequences and predict future candles autoregressively. They are designed as a unified base for downstream quantitative tasks such as price forecasting, volatility prediction, and regime detection.

Model family

Model	Tokenizer	Context	Params
Mhondoro-mini	Mhondoro-Tokenizer-2k	2048	4.1M
Mhondoro-small	Mhondoro-Tokenizer-base	512	24.7M
Mhondoro-base	Mhondoro-Tokenizer-base	512	102.3M

Quick start

from model.mhondoro import Mhondoro, MhondoroTokenizer, MhondoroPredictor

tokenizer = MhondoroTokenizer.from_pretrained("chifural/Mhondoro-Tokenizer-base")
model = Mhondoro.from_pretrained("chifural/Mhondoro-small")
predictor = MhondoroPredictor(model=model, tokenizer=tokenizer, device="cuda:0")

forecast = predictor.predict(
    df=ohlcv_dataframe,
    x_timestamp=context_timestamps,
    y_timestamp=forecast_timestamps,
    pred_len=120,
)

Source code: github.com/vigilancetrent/mhondoro.

License

MIT. Copyright (c) 2026 Louis Chifura.

Notice

This work builds on the architectural design of the open-source NeoQuasar/Kronos project (MIT, Copyright (c) 2025 ShiYu). The Mhondoro fork extends the original with additional fine-tuning workflows, broader market coverage, and an evolving feature set. See the upstream project at https://github.com/Shiyu-Coder/Kronos for the historical record.