Recently, Google DeepMind and Google Research officially released WeatherNext 2, a next‑generation AI weather forecasting model capable of generating high‑resolution predictions up to eight times faster and extending forecast coverage to the 0–15 day range. Google's official blog and DeepMind's research page published the model's technical capabilities today and confirmed that WeatherNext 2 has been integrated into Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API, offering finer‑grained weather experiences to users worldwide.
DeepMind states WeatherNext 2 is its most accurate AI weather model to date, outperforming prior systems on most meteorological variables. The model learns from multi‑source data, including satellite imagery, to generate forecasts across different time scales. Google emphasizes that the model improves both speed and resolution, helping users obtain timelier information for daily life, travel, and risk management. The DeepMind research team said, "WeatherNext 2 represents the latest stage in AI weather forecasting. It complements and enhances existing forecasting systems with greater speed and broader temporal coverage."
Official information indicates WeatherNext 2 outputs are available to developers via Google Cloud Vertex AI, accompanied by datasets for geoscience researchers, enterprises, and development teams to analyze. The model's core innovation is the Functional Generative Network (FGN) architecture, which injects perturbation signals to directly generate hundreds of possible weather scenarios, ensuring physical realism and diversity in forecasts. This approach allows the model to complete computations on a single TPU in under a minute, whereas traditional physics‑based simulations can take hours. DeepMind's blog explains that WeatherNext 2 outperforms its predecessor GenCast on 99.9% of variable and forecast‑window combinations (such as temperature, wind speed, humidity), with an average Continuous Ranked Probability Score (CRPS) improvement of 6.5%, and notable gains in ensemble mean RMSE and economic value for extreme events.
Over recent years, DeepMind has advanced AI applications in weather prediction. Since the 2021 introduction of the GraphCast model, these AI systems have been gradually integrated into Google products to support forecasts for hundreds of millions of users. WeatherNext 2 continues this trajectory, further extending forecast range and operational capability. DeepMind's collaboration with the European Centre for Medium‑Range Weather Forecasts (ECMWF) is a key part of the effort: the model produces four 6‑hourly forecasts per day and supports timely hour‑level decision‑making. Bloomberg reported that WeatherNext 2, combined with NVIDIA H100 GPU clusters, can deliver hourly forecasts in minutes compared with hours on traditional supercomputers. DeepMind research scientist Ferran Alet said in an interview, "It provides more granular forecasts to help industries make precise decisions." For example, in European tests the model achieved up to 97.2% accuracy on certain targets for variables such as wind speed, precipitation, and pressure, outperforming the ECMWF baseline by up to 20% in some cases.
WeatherNext 2 has moved from the lab to production use. DeepMind says the model is embedded in the Google Cloud Weather API, allowing enterprise customers to customize forecasts, such as routing logistics to avoid storms. Analysis by MarkTechPost notes that FGN tools can simulate mesoscale atmospheric dynamics to support disaster warning—such as projecting tropical cyclone track, intensity, and structure and simulating landfall up to 15 days in advance. This capability is critical in the climate crisis era: during the 2025 hurricane season, similar models reportedly helped regions like Florida optimize evacuation routes, potentially saving hundreds of millions of dollars. An ECMWF report suggests AI weather models like WeatherNext 2 could reduce global economic losses by 15%. The model is also adept at predicting rare events such as extreme rainfall or snowfall, although DeepMind acknowledges training data gaps could cause about 10% error and pledges continued improvement.
Despite many details remaining unpublished, DeepMind stresses that AI weather forecasting should be used in combination with traditional meteorological models to ensure accuracy, safety, and interpretability. The model is currently in early access for enterprises and developers and will be iteratively improved based on usage. DeepMind research scientist Akib Uddin said in a YouTube interview, "It is not a replacement for human experts but a powerful ally." For example, when a user queries "probability of heavy rain in California this weekend," the model can produce personalized forecasts that include risk maps and alternate routes. The official blog emphasizes that the model's generative approach maintains physical realism via perturbation signals and supports diverse scenarios from global to regional scales.
The release has far‑reaching implications. Traditional forecasting relies on physical equations and supercomputers, which are costly and update slowly. WeatherNext 2's AI approach lowers the barrier: training data come from public satellite sources (such as NOAA), and computational demand is reportedly reduced by 50%. In the climate era this could mean more accurate disaster responses. DeepMind's collaboration with ECMWF has already impacted a billion users, and Google Cloud Next 2025 previewed integration with the Vertex AI platform with an enterprise subscription price reportedly at $0.01 per forecast. NVIDIA's developer blog notes that combined with Earth‑2, WeatherNext 2 could simulate forecasts beyond two weeks, helping shift climate action from reactive to predictive.
Challenges remain. DeepMind acknowledges the model still has errors in predicting extreme events such as volcanic ash clouds and needs more historical data for training. Privacy concerns arise because satellite data can involve sensitive locations, and EU GDPR requires audit compliance. Uddin warned, "AI forecasts are powerful but require human oversight—the black‑box nature of some models can amplify uncertainties." DeepMind responded that access will be provided through Earth Engine and BigQuery to enable global researchers to contribute data. WeatherNext 2 is currently limited to early access for enterprises and developers, with public service expansion expected in 2026.
From a broader perspective, WeatherNext 2 strengthens DeepMind's leadership in climate AI. Since partnering with ECMWF in 2023, DeepMind has released three weather models affecting hundreds of millions of users. The official blog reiterates that the model's generative method preserves physical realism via perturbation signals and supports scenarios from global to regional scales. Engadget reported strong performance on key variables such as wind speed, precipitation, and pressure, and noted that WeatherNext 2 is available to developers through Google Cloud Vertex AI in early access. Bloomberg highlighted the model's potential in energy trading, where it can help traders simulate wind and solar radiation changes to optimize grid stability.
The launch of WeatherNext 2 may usher in a new era of AI‑driven weather forecasting. Ahead of COP30, it is a reminder that technology is not omnipotent but can amplify human judgment. DeepMind expects the model to cover major global cities by 2026. Silicon Valley watches closely: how will AI descend from the cloud to the ground?