Back home

TimeH2O-v1

x 0.62
Visit source

Intro

TimeH2O-v1 is a time-series foundation model released by H2O.ai on October 26, 2026. With about 108 million parameters, it forecasts zero-shot: it reads a series' history and predicts the future without being trained or fine-tuned on that data.

At launch it ranked #1 of 32 models on the fev-bench leaderboard (100 real-world forecasting tasks) by win rate (91.9%) and skill score (50.8), ahead of Mistral's Timestral-209M, t0-gamma, Google's TimesFM-3 and Amazon's Chronos-2. It also posts the lowest aggregate MASE (0.654) and CRPS (0.441) among zero-shot models on GIFT-Eval. The weights are on Hugging Face under the H2O Model Non-Commercial License, alongside a 5.8M-parameter TimeH2O-v1-tiny; H2O.ai kept promoting it on X through Halloween ("Your forecasting model shouldn't take longer to train than it takes to become a skeleton").

Explain

Model details

  • Size: ~108M parameters (TimeH2O-v1); TimeH2O-v1-tiny has 5.8M parameters for maximum throughput.
  • Context: up to 16,384 time steps; horizon is unlimited.
  • Output: probabilistic forecasts as 9 quantiles (0.1–0.9).
  • Inputs: univariate target series, with optional covariates (price, promotions, weather, etc.) — the tiny model does not support covariates.
  • Architecture: a purpose-built transformer with a flip-equivariant design (a flipped series gets the flipped forecast), trained on real-world series plus synthetic data generated with temporal causal models.

Benchmarks (as of Oct 23–24, 2026)

Benchmark Metric TimeH2O-v1 Timestral-209M TimesFM-3 Chronos-2
fev-bench Skill score ↑ 50.8 50.4 49.0 47.3
GIFT-Eval MASE ↓ 0.654 0.658 0.667 0.698
GIFT-Eval CRPS ↓ 0.441 0.446 0.456 0.485

TimeH2O-v1 was never trained on any fev-bench dataset.

Licensing

The weights are open but non-commercial: internal benchmarking and testing are allowed, while production use requires a commercial license from H2O.ai. The h2ofm inference code is Apache-2.0. Access on Hugging Face is gated (accept the terms first).

Why it matters

Time-series foundation models let companies replace hundreds of per-series models (demand, inventory, cash, energy load, patient volume) with one pretrained model. H2O.ai positions TimeH2O as the forecasting layer that AI agents can call for grounded numbers.

Usage

Quick start

  1. Request access at https://huggingface.co/h2oai/timeh2o-v1 (accept the non-commercial license).
  2. Install H2O's inference package (h2ofm, Apache-2.0) as described on the model card.
  3. Load and forecast:
import numpy as np
from h2ofm import TimeH2O  # see the model card for the exact install/import

model = TimeH2O.from_pretrained("h2oai/timeh2o-v1")
forecast = model.forecast(np.sin(np.arange(200) / 10), horizon=24)
forecast.quantiles  # shape (24, 9): quantiles 0.1 … 0.9
  1. Pass covariates (e.g. price or weather) when you have them, to improve accuracy.
  2. Use the 0.1/0.9 quantiles as an uncertainty band for planning, not just the median.

Good fits

  • Demand, inventory and sales forecasting with many short series.
  • Energy load, web traffic and operations metrics.
  • Quick baselines before investing in per-series models.

For commercial deployment, contact H2O.ai for a license; you can also fine-tune on your own data and infrastructure.

Links