WeatherNext 2 AI brings faster, more accurate forecasts to Google
Google has introduced WeatherNext 2, an AI model delivering faster, higher‑resolution forecasts with probabilistic outputs. Data is available in Earth Engine and BigQuery, with an early access programme on Vertex AI for custom inference.
Google has unveiled WeatherNext 2, its most advanced AI weather forecasting model to date, bringing higher‑resolution and more efficient predictions to products including Search, Gemini, Pixel Weather and Google Maps.
Announced on 17 November 2025, the rollout marks a move from experimental research to mainstream deployment.
WeatherNext 2 is a global, medium‑range forecasting system developed by Google DeepMind and Google Research.
The model has delivered forecasts up to eight times faster than its predecessor and has surpassed the previous WeatherNext generation on 99.9% of tested variables and lead times, according to Google.
It can produce hundreds of plausible weather scenarios from a single starting point, aiding risk‑aware planning.
How does the new approach work?
At the core of WeatherNext 2 is a Functional Generative Network (FGN) that injects controlled randomness into the model’s architecture, allowing it to generate many coherent outcomes in one run.
On Google’s TPU hardware, a forecast can be generated in under a minute — a task that conventional physics‑based models typically complete in hours.
Google said WeatherNext 2 has improved temporal resolution (down to hourly for selected fields) and broader accuracy gains across variables such as temperature, wind and humidity, and across lead times out to 15 days.
Google made WeatherNext 2 forecast data available in Earth Engine and BigQuery for analysis, with an early access programme on Google Cloud’s Vertex AI for custom, on‑demand inference. This access has extended the model beyond pre‑computed datasets, enabling users to configure runs to their own needs.
Model specifications at a glance
- Architecture: Functional Generative Network; global coverage at 0.25° spatial resolution.
- Time steps: hourly for selected variables (others at six‑hourly), with forecast initialisations every 00, 06, 12 and 18 UTC.
- Forecast horizon: up to 15 days; default ensemble output of 64 members (four sub‑models, 16 members each).
- Data and inputs: trained on ERA5 and ECMWF HRES datasets; initialised from HRES‑fc0; historical backtesting from January 2024 available.
Rapid, probabilistic forecasts have been sought by sectors from energy and agriculture to transport and logistics.
By integrating WeatherNext 2 into core Google services and opening pathways for researchers and enterprises to run custom inference, Google has positioned AI‑based forecasting alongside traditional numerical models at consumer and operational scale.


