Table of Contents
Overview
In the rapidly evolving landscape of artificial intelligence, Google DeepMind and Google Research continue to push boundaries, and their latest achievement, WeatherNext 2, represents a significant advancement in weather forecasting technology. Launched in November 2025, this isn’t merely another weather prediction system; it’s Google DeepMind’s most advanced and efficient AI weather forecasting model family, engineered to deliver forecasts up to 8 times faster than its predecessor with unprecedented hourly resolution capabilities. By leveraging a groundbreaking Functional Generative Network architecture, WeatherNext 2 generates hundreds of realistic probabilistic scenarios, providing nuanced understanding of uncertainty and potential for extreme events. Now powering upgraded forecasts across critical Google services including Search, Gemini, Pixel Weather, and Google Maps Platform’s Weather API, and accessible through Earth Engine, BigQuery, and Vertex AI, WeatherNext 2 is positioned to become the new standard in AI-driven weather intelligence.
Key Features
WeatherNext 2 delivers sophisticated capabilities designed for accuracy, speed, and comprehensive meteorological insight across medium-range forecasting.
8x Faster Forecast Generation: Experience significantly accelerated forecast delivery compared to the previous WeatherNext model, with each prediction taking less than one minute on a single Tensor Processing Unit (TPU), enabling more timely decision-making.
Hourly Resolution Forecasts: Receive highly detailed predictions broken down by the hour (1-hour resolution), offering granular insights into rapidly changing weather conditions, a significant improvement from previous 6-hour timesteps.
Functional Generative Network Architecture: At its core, this innovative AI architecture allows the model to create diverse and physically realistic weather scenarios by injecting noise directly into the model structure, moving beyond single-point deterministic predictions.
Hundreds of Probabilistic Scenarios: Instead of generating a single forecast, WeatherNext 2 provides a range of possible outcomes from each starting point, helping users understand the likelihood and potential impact of various weather events including worst-case scenarios.
Improved Extreme Event Prediction: The generative nature of the FGN model enhances its ability to anticipate and quantify uncertainty around severe weather, from tropical cyclones and hurricanes to heatwaves and atmospheric rivers, with approximately one extra day of useful predictive skill for cyclone tracking compared to previous models.
15-Day Forecast Horizon: Provides global medium-range forecasts up to 15 days in advance with 0.25-degree latitude-longitude resolution (approximately 28km x 28km at the equator), covering 6 atmospheric variables at 13 pressure levels plus 6 surface variables.
Integration with Google Search, Gemini, Pixel Weather, and Maps: Seamlessly access WeatherNext 2’s advanced forecasts directly within the Google products used daily by billions of users worldwide, democratizing access to cutting-edge weather intelligence.
Available via Earth Engine, BigQuery, and Vertex AI: For developers, researchers, and enterprises, the model’s capabilities are accessible through Google’s powerful cloud platforms, enabling custom applications, advanced analytics, and downstream model training.
Joint Forecasting Capabilities: Beyond predicting individual variables in isolation, the model skillfully forecasts interconnected weather systems such as entire regions affected by high heat or expected power output across wind farms, understanding how individual meteorological pieces fit together.
How It Works
Understanding WeatherNext 2’s sophisticated architecture reveals Google DeepMind’s breakthrough approach to probabilistic weather forecasting.
The process begins with an advanced AI model rigorously trained on the ERA5 reanalysis dataset, encompassing decades of atmospheric physics data and extensive historical weather observations from 1979 to 2018. This foundational training allows the model to grasp the complex dynamics of Earth’s atmosphere and learn cause-and-effect relationships governing weather evolution.
When a forecast is requested, the system processes current atmospheric conditions, feeding this real-time data into its innovative Functional Generative Network. The FGN architecture uses independently trained neural networks and injects noise in function space to create coherent variability in weather forecast predictions. This approach differs fundamentally from previous diffusion models by generating realistic joint distributions of interconnected variables.
The network then generates an ensemble of forecasts, producing hundreds of probabilistic scenarios rather than a single deterministic outcome, each taking less than one minute on a single TPU chip. These diverse scenarios capture the full range of possibilities including low-probability but catastrophic events. The model iteratively forecasts forward in 6-hour timesteps up to 15 days, while providing hourly granularity within each period.
Finally, these highly detailed and rapid forecasts are delivered directly to users through various Google products and made available for advanced use via Google’s cloud platforms. The system generates forecasts four times daily, providing timely data to support better decision-making across industries.
Use Cases
The versatility of WeatherNext 2 extends across numerous industries and applications, offering critical insights for diverse meteorological needs.
Agriculture and Farming: Optimize planting, irrigation, and harvesting schedules by anticipating weather patterns with hourly precision, maximizing crop yields while minimizing resource waste.
Disaster Preparedness and Emergency Response: Enhance readiness and response strategies for severe weather events including tropical cyclones, floods, and extreme heat by leveraging probabilistic forecasts and improved extreme event prediction with extended lead times.
Aviation and Transportation Planning: Improve flight scheduling, route optimization, and ground transportation logistics by accounting for detailed, fast-updating weather information with hourly resolution.
Energy Demand Forecasting and Renewable Energy: Better predict energy consumption spikes or dips related to temperature changes, aiding in grid management and resource allocation. Forecast wind power generation and solar potential with greater accuracy for renewable energy planning.
Outdoor Event Planning: Make informed decisions for concerts, sports events, and festivals by understanding hourly weather changes and potential extreme conditions with greater confidence.
Climate Research and Scientific Analysis: Utilize the model’s advanced capabilities and data access via Earth Engine and BigQuery for in-depth studies on climate patterns, extreme weather trends, and atmospheric dynamics.
Global Supply Chain Management: Anticipate weather-related disruptions to logistics, shipping, and transportation networks with extended forecast horizons, enabling proactive contingency planning.
Search Engine Weather Integration: Provide Google users with the most accurate and up-to-date weather information directly within search results, powered by state-of-the-art AI forecasting.
Mobile Weather Applications: Power next-generation weather apps through Pixel Weather integration with faster, more detailed, and probabilistically rich forecasts accessible to millions.
Financial Trading and Risk Management: Enable energy traders, commodity markets, and insurance companies to make more informed decisions based on accurate weather predictions with hourly granularity.
Water Resource Management: Support reservoir operations, flood control, and drought planning with improved precipitation forecasts and extreme event detection.
Pricing and Availability
WeatherNext 2 is accessible through multiple channels designed to serve different user needs from individual consumers to enterprise developers.
Consumer Access: Free integration into Google Search, Gemini AI assistant, Pixel Weather app, and Google Maps provides billions of users worldwide with access to WeatherNext 2 powered forecasts at no cost.
Google Maps Platform Weather API: Available for businesses integrating weather data into applications through Google Maps Platform with standard API pricing.
Earth Engine and BigQuery: WeatherNext 2 forecast data now available in Google Earth Engine and BigQuery for researchers and data scientists, following standard pricing for these Google Cloud services.
Vertex AI Early Access Program: Custom model inference available through early access program on Google Cloud’s Vertex AI platform, enabling businesses to customize WeatherNext 2 models for specific use cases and regions with enterprise-grade cloud infrastructure.
Open Research Access: Following Google DeepMind’s pattern with previous weather models (GraphCast and GenCast code already open-sourced), WeatherNext 2 model code, weights, and forecasts are expected to be released to support the wider weather forecasting community, enabling researchers and meteorologists globally to benefit from and build upon this technology.
Pros and Cons
Understanding the capabilities and considerations of WeatherNext 2 provides clarity for assessing its value across different applications.
Advantages
8x Faster Than Previous Model: Delivers forecasts with unprecedented speed compared to WeatherNext 1, taking under one minute on single TPU versus hours on supercomputers for traditional physics-based ensemble forecasts, crucial for time-sensitive decisions.
Hourly Resolution Detail: Provides 1-hour granular insights into weather changes, offering a level of temporal detail previously difficult to achieve with most operational systems using 6-hour or 12-hour timesteps.
Probabilistic Uncertainty Quantification: Moves beyond single deterministic predictions to offer hundreds of scenarios from each starting point, giving users clearer picture of potential outcomes, risks, and low-probability but high-impact events.
Superior Performance on 99.9% of Variables: Outperforms previous state-of-the-art WeatherNext model on 99.9% of variables (temperature, wind, humidity, pressure) and lead times (0-15 days) with average CRPS improvement around 6.5% and maximum gains near 18% for some variables.
Enhanced Extreme Event Prediction: Demonstrates improved tropical cyclone tracking with approximately one extra day of useful predictive skill, better atmospheric river identification, and more accurate extreme temperature forecasting.
Widespread Google Ecosystem Integration: Seamlessly available across popular Google products reaching billions of users, enhancing accessibility and democratizing advanced weather intelligence beyond specialist communities.
Joint Distribution Forecasting: Skillfully predicts interconnected weather systems rather than just independent variables, enabling useful predictions like regions affected by high heat or wind farm power output.
Cloud Platform Accessibility: Available through Earth Engine, BigQuery, and Vertex AI for developers and enterprises to build custom applications and integrate into decision-making workflows.
Disadvantages
Proprietary Google Technology: Being a Google DeepMind and Google Research product, core technology details and full model architecture may not be immediately transparent, though following open-source release pattern of previous models.
Dependent on Google Infrastructure: Relies heavily on Google’s cloud and product ecosystem, potentially limiting standalone deployment options for organizations preferring platform-agnostic solutions.
Cloud Platform Access Required for API Use: Developers and enterprises needing API access beyond consumer products will integrate with Google Cloud platforms like Earth Engine, BigQuery, or Vertex AI, incurring associated costs and requiring cloud expertise.
Limited to Weather Forecasting Domain: While exceptionally powerful in meteorological applications, its specialized focus means it doesn’t extend to other AI domains or broader climate modeling tasks.
Challenges with Extreme Precipitation: Researchers acknowledge the model still faces difficulties predicting extreme torrential rain and blizzards, mainly because relevant observations in training data are sparse for these rare events.
Early Access for Custom Inference: Vertex AI custom model inference currently in early access program, meaning full production availability and pricing may still be evolving for enterprise customization use cases.
How Does It Compare?
The weather forecasting landscape in 2025 encompasses both traditional numerical weather prediction systems and emerging AI-powered models. Understanding WeatherNext 2’s position requires examining specific competitors across both categories.
Google DeepMind Weather Model Family
WeatherNext 2 represents the latest evolution in Google’s AI weather forecasting suite, which includes GraphCast (2023), GenCast (2024), NeuralGCM, MetNet-3, and flood/wildfire prediction models. GraphCast pioneered 10-day deterministic forecasts with Graph Neural Networks, achieving 90% accuracy advantage over ECMWF HRES. GenCast advanced to 15-day probabilistic ensemble forecasts using diffusion models, outperforming ECMWF ENS on 97.2% of targets.
WeatherNext 2 builds upon GenCast with its Functional Generative Network architecture, achieving 8x faster generation, 1-hour resolution capability, and improved performance on 99.9% of variables compared to WeatherNext Gen (GenCast-based). The FGN model contains approximately 180 million parameters per model seed versus GenCast’s 57 million, with latent dimension 768 and 24 transformer layers versus GenCast’s 512 and 16 layers, representing substantial architectural advancement.
ECMWF (European Centre for Medium-Range Weather Forecasts)
ECMWF operates the gold-standard HRES deterministic system and ENS ensemble forecast system, widely adopted globally. ECMWF recently launched its own AI-powered forecasting system (AIFS) in February 2025, marking traditional weather agencies’ embrace of machine learning. ECMWF systems rely on physics-based numerical weather prediction using sophisticated general circulation models running on supercomputers.
WeatherNext 2 differentiates through dramatically faster computation (under 1 minute on single TPU versus hours on supercomputers with tens of thousands of processors), probabilistic ensemble generation at scale, and demonstrated accuracy improvements in head-to-head comparisons. However, ECMWF maintains advantages in institutional trust, decades of operational experience, and comprehensive data assimilation capabilities. The organizations increasingly collaborate, with ECMWF experimentally running Google’s weather models on its website.
NOAA GFS (Global Forecast System)
The US National Oceanic and Atmospheric Administration’s GFS provides free, global weather forecasts up to 16 days using physics-based numerical weather prediction. GFS runs four times daily on NOAA supercomputers and serves as backbone for many US weather services and commercial forecast providers.
WeatherNext 2 offers comparable forecast horizon with dramatically reduced computational requirements and faster updates. While GFS benefits from extensive observational data assimilation and operational validation over decades, WeatherNext 2’s AI approach enables probabilistic uncertainty quantification and demonstrated superior accuracy on multiple verification targets. GFS’s free global availability and established institutional use remain advantages for organizations preferring traditional NWP methods.
Pangu-Weather (Huawei)
Pangu-Weather, developed by Huawei Cloud, represents China’s entry into AI weather forecasting with deterministic 10-day global forecasts using three-dimensional neural network architecture. It demonstrated competitive accuracy with ECMWF HRES while running significantly faster.
WeatherNext 2 distinguishes itself through probabilistic ensemble forecasting versus Pangu-Weather’s deterministic approach. The hundreds of scenarios generated by WeatherNext 2’s FGN provide richer uncertainty quantification essential for risk management and extreme event planning. WeatherNext 2’s 1-hour resolution and 15-day horizon also extend beyond Pangu-Weather’s capabilities, while Google’s ecosystem integration provides broader accessibility.
FourCastNet (NVIDIA)
FourCastNet, developed by NVIDIA and collaborators, uses Adaptive Fourier Neural Operators for rapid data-driven weather prediction. It emphasizes extreme weather detection and runs efficiently on GPUs with global forecasts in seconds.
Both systems prioritize speed and extreme event prediction, but WeatherNext 2’s FGN architecture provides more comprehensive probabilistic ensemble generation. WeatherNext 2’s production integration across Google services also provides real-world validation at massive scale that research models lack. However, FourCastNet’s GPU optimization may offer deployment flexibility for organizations with existing NVIDIA infrastructure.
IBM Weather (The Weather Company)
IBM Weather, through The Weather Company, provides commercial weather forecasting services combining traditional NWP models, proprietary data sources, and increasingly AI-enhanced prediction. It serves enterprise customers across industries with customized forecasting solutions.
WeatherNext 2 offers technological advantages in speed and probabilistic modeling but currently operates within Google’s ecosystem. IBM Weather’s strength lies in comprehensive commercial service offerings, established enterprise relationships, and multi-source data integration. For organizations seeking turnkey commercial weather services with dedicated support, IBM Weather provides proven alternative, while WeatherNext 2 excels in cutting-edge AI capabilities and consumer reach.
AccuWeather
AccuWeather operates as a major commercial weather service providing consumer forecasts, enterprise solutions, and media content. It combines multiple NWP model outputs with proprietary forecasting techniques and human meteorologists’ expertise.
WeatherNext 2’s pure AI approach represents fundamentally different methodology from AccuWeather’s hybrid human-AI system. While AccuWeather benefits from brand recognition, established consumer base, and meteorologist interpretation for complex situations, WeatherNext 2 offers superior computational efficiency and demonstrated accuracy on standardized benchmarks. The integration into Google services also provides distribution advantage reaching billions of users.
Microsoft Azure Weather Services
Microsoft provides weather data and forecasting through Azure Maps and partnerships with weather data providers. Azure offers APIs for integrating weather information into applications with global coverage.
WeatherNext 2 and Azure Weather Services serve similar developer needs but with different underlying approaches. WeatherNext 2’s advanced AI forecasting and probabilistic scenarios represent technical advantages, while Azure’s multi-cloud strategy and enterprise Microsoft ecosystem integration appeal to organizations preferring Microsoft technology stack. Pricing models and service level agreements differ significantly between Google Cloud and Azure platforms.
ECMWF AIFS (AI Forecasting System)
ECMWF launched its AI Forecasting System in February 2025, marking traditional meteorological agency’s adoption of machine learning. AIFS combines ECMWF’s data assimilation expertise with neural network prediction, aiming to complement rather than replace physics-based systems.
This represents important validation of AI weather forecasting approach. WeatherNext 2 and ECMWF AIFS will likely coexist and potentially collaborate, with WeatherNext 2 offering advantages in computational efficiency and Google ecosystem integration, while ECMWF AIFS benefits from institutional authority, operational meteorological expertise, and integration with traditional forecasting workflows trusted by national weather services globally.
Key Differentiators
WeatherNext 2’s unique positioning centers on several breakthrough capabilities. The 8x speed advantage over its predecessor combined with sub-minute generation time on single TPU democratizes access to ensemble forecasting previously requiring massive supercomputer resources. The Functional Generative Network architecture generating hundreds of physically realistic, interconnected scenarios provides richer uncertainty quantification than deterministic or limited-ensemble approaches.
The 1-hour resolution capability offers temporal granularity particularly valuable for energy trading, transportation, and event planning where hourly differences significantly impact decisions. Performance improvements on 99.9% of variables with average 6.5% CRPS gain and up to 18% maximum improvement demonstrate consistent advancement across the entire meteorological variable space.
Most significantly, production integration across Google Search, Maps, Gemini, and Pixel Weather reaching billions of daily users represents unprecedented scale for AI weather forecasting deployment. This moves cutting-edge meteorological AI from research papers and specialized applications into everyday digital experiences, fundamentally changing public access to advanced weather intelligence.
For pure computational efficiency and AI innovation, WeatherNext 2 leads the field. For institutional trust and operational meteorology integration, ECMWF systems remain authoritative. For commercial services with comprehensive support, IBM Weather and AccuWeather offer proven solutions. For organizations seeking state-of-the-art AI forecasting within Google Cloud ecosystem with pathways to customization through Vertex AI, WeatherNext 2 presents the most advanced option available.
Final Thoughts
WeatherNext 2 represents a significant milestone in AI-driven weather forecasting, combining breakthrough speed, granular hourly resolution, and sophisticated probabilistic insights powered by Google DeepMind’s innovative Functional Generative Network architecture. Its demonstrated performance improvements across 99.9% of meteorological variables, enhanced extreme event prediction, and 8x faster generation compared to its predecessor position it at the forefront of weather intelligence technology.
The widespread integration across Google Search, Gemini, Pixel Weather, and Google Maps democratizes access to cutting-edge forecasting capabilities, reaching billions of users globally. For researchers and enterprises, availability through Earth Engine, BigQuery, and Vertex AI provides pathways for custom applications and advanced analytics, though requiring Google Cloud platform adoption.
While considerations include reliance on Google infrastructure, early-stage custom inference access, and acknowledged challenges with extreme precipitation events, the model’s ability to generate hundreds of probabilistic scenarios in under a minute fundamentally changes what’s possible in operational weather forecasting. Whether for critical enterprise decisions requiring probabilistic risk assessment, scientific research advancing atmospheric understanding, energy trading demanding hourly precision, or simply planning daily activities, WeatherNext 2 delivers a new era of intelligent, accessible, and actionable weather prediction that empowers users with unprecedented clarity and foresight in our changing climate.