H2O.ai’s TimeH2O Beats Mistral, Google, and Amazon in Time-Series Forecasting
The #1 open-weight time-series pre-trained foundation model leads fev-bench’s 100 tasks, a sign that purpose-built predictive models will ground the forecasts AI agents act on.
H2O.ai, the trusted Sovereign AI Platform with the software and expertise to run predictive, generative, and agentic AI in production, today released TimeH2O-v1, a time-series foundation model that ranks first of 32 models on the fev-bench leaderboard by win rate. TimeH2O-v1 posted a 91.9% win rate and a skill score of 50.8 across the benchmark’s 100 forecasting tasks, ahead of Timestral-209M, t0-gamma, TimesFM-3, and Chronos-2.
Read more – Check Point Launches Firewall-as-a-Subscription
“Building your business with accurate forecasting and prediction is the superpower of AI,” said Sri Ambati, Founder and CEO of H2O.ai
Forecasts run the business: demand, inventory, cash, energy load, patient volume. Most organizations still build and maintain a separate model for each of those series, and retrain it when the data shifts. A forecasting foundation model works zero-shot: it reads a series’ history and forecasts it with no training on that data. TimeH2O-v1 was never trained on any fev-bench dataset, and it still ranks first.
Read more –Datalink Networks Addresses Critical Global Cybersecurity Risks
“Building your business with accurate forecasting and prediction is the superpower of AI. TimeH2O gives organizations a generalized, pre-trained model for time series that can then be fine-tuned into a domain-specific foundation model for each customer’s use case,” said Sri Ambati, Founder and CEO of H2O.ai.
Read more –Cyera Achieves FedRAMP High Status to Enable Federal AI
How TimeH2O-v1 performs
fev-bench scores models zero-shot on real time series across many domains and frequencies, measuring probabilistic accuracy against a seasonal-naive baseline. As of October 23, 2026, on the leaderboard’s default view:
- First by win rate. TimeH2O-v1 beats the other models on 91.9% of head-to-head comparisons, ranking first of 32. Timestral-209M follows at 89.6% and t0-gamma at 88.1%.
- A skill score of 50.8. Its forecast error is 50.8% lower than the seasonal-naive baseline’s, aggregated across all 100 tasks.
- Zero-shot. At forecast time, the model reads each series’ history and predicts from it. No training or fine-tuning happens: its weights never change.
Write to us editorial@demandmediaagency.com. to learn more about our exclusive editorial packages and programmes.
