You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

TimeH2O-v1

TimeH2O-v1 is a 108M-parameter time series foundation model for zero-shot probabilistic forecasting, with support for covariates.

📰 Blog · 💻 GitHub (coming soon) · 📄 Paper (coming soon)

(a) Architecture · (b) Training batch · (c) Transformer block.

Benchmark Results

TimeH2O-v1 achieves state-of-the-art zero-shot results on GIFT-Eval and fev-bench.

Benchmark	Metric	TimeH2O-v1	Timestral-209M	t0-gamma	TimesFM-3	Chronos-2
GIFT-Eval	CRPS ↓	0.441	0.446	0.452	0.456	0.485
	MASE ↓	0.654	0.658	0.661	0.667	0.698
fev-bench	SQL skill score ↑	50.83	50.39	49.42	48.99	47.28
	MASE skill score ↑	39.95	39.49	37.99	37.82	35.50

TimeH2O family

Model	Parameters	Covariates	Best for
TimeH2O-v1	108M	✅	Max accuracy
TimeH2O-v1-tiny	5.8M	❌	Max throughput

Model Details

Context: up to 16,384 time steps
Horizon: unlimited
Output: 9 quantiles (0.1 to 0.9)
Inputs: target series with optional covariates

Usage

pip install h2ofm

import numpy as np
from h2ofm import TimeH2O

model = TimeH2O.from_pretrained("h2oai/timeh2o-v1")
forecast = model.forecast(np.sin(np.arange(200) / 10), horizon=24)

forecast.point  # (24,), median forecast; also forecast.median
forecast.quantiles  # (24, 9), default levels 0.1 through 0.9
forecast.quantile(0.9)  # (24,)