Today, Google DeepMind and Google Research released WeatherNext 3, a new artificial intelligence model that marks a significant leap in the ability to forecast our planet’s chaotic atmosphere. Promising sharper resolution, hourly updates, and direct integration into Google products like Search, Maps, and Gemini, the model is already being hailed as the most accurate among leading contenders tested on the Operational WeatherBench, a benchmark built by startup Brightband. This isn’t just another incremental update; WeatherNext 3 directly confronts the long-standing weaknesses of AI weather forecasting—coarse resolution, poor rainfall prediction, and dependence on preformatted government data—offering a future where weather information is more granular, frequent, and actionable.
How WeatherNext 3 Outperforms Both AI Rivals and Traditional Models
The model has proven its mettle against a crowded field. According to Brightband’s benchmark, WeatherNext 3 beats out other deep-learning systems from Google itself, Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts (ECMWF). More strikingly, it also surpasses the traditional physics-based forecasts produced by the US National Weather Service and the ECMWF, which have been the gold standard for decades. The benchmark evaluates metrics like temperature, windspeed, and humidity, providing a clear picture of where each model stands.
Traditional weather forecasting relies on supercomputers solving complex mathematical equations that describe atmospheric physics. These systems are remarkably accurate but expensive and slow to run. The ECMWF’s decision to release over half a century of its reanalysis data in 2018 opened the door for deep learning. Researchers trained models to make predictions far more quickly, achieving comparable accuracy to government tools. Now, WeatherNext 3 pushes that frontier further, specifically targeting the three main weaknesses of first- and second-generation AI models.
Unprecedented Resolution: From 25 km Down to 5 km
One of the biggest criticisms of AI weather models has been their coarse spatial resolution. Earlier models typically forecast over areas as large as 15 to 25 square kilometers, which is too broad for localized decision-making. WeatherNext 3 tackles this head-on. On key surface variables, researchers told TechCrunch that the model can predict conditions down to a resolution of approximately 5 kilometers. This dramatic improvement brings AI forecasts closer to the granularity needed for agriculture, aviation, and emergency planning, moving beyond generalized regional outlooks to city- or even neighborhood-level insights.
Rain Prediction Improves by 60 Percent Over WeatherNext 2
Precipitation has historically been a weak spot for machine learning models in meteorology. The chaotic nature of rainfall makes it notoriously difficult to predict using pattern recognition alone. WeatherNext 3 shows a 60 percent improvement in rainfall evaluations compared to its predecessor, WeatherNext 2. This is not a minor tweak; it represents a fundamental advance in the model’s ability to handle one of the most impactful weather variables. For industries like agriculture, logistics, and insurance, better rain forecasts translate directly into economic value.
Hourly Forecasts Replace the Standard Six-Hour Cycle
Most operational weather models produce forecasts every six hours. WeatherNext 3 shifts to an hourly cadence, dramatically increasing the temporal density of predictions. This is made possible because the model can now ingest raw satellite data collected in real-time on an hourly basis. Typically, AI models rely on the analysis products generated by supercomputers, but feeding on raw empirical observations promises a more accurate and timely forecast. Google claims that WeatherNext 3 is the first AI model to directly incorporate raw observations for a high-resolution global forecast, though AI startup WindBorne has been doing something similar with its WeatherMesh 6 model since late 2025, using its own fleet of weather balloons. Google counters that its forecasts are higher resolution across the entire globe. Regardless of the order, both models still depend on national weather datasets to perform forecasts, so true direct data assimilation remains a work in progress.
What Is the WeatherNext 3 Architecture and Why Does It Matter?
The improvements in WeatherNext 3 stem from deliberate architectural choices. The model is significantly larger, with 2.4 times more parameters than its predecessor. This increased capacity allows it to learn more complex patterns from the training data. More importantly, the designers tailored the decoder heads to produce more useful outputs. While most weather models output metrics averaged across a 3D grid, DeepMind researchers tuned this model to also visualize specific phenomena like cyclone paths, a feature that has already won plaudits in the meteorological community.
The model is also trained to target its forecasts to specific weather data stations. This is a critical innovation. By predicting what a given station—say, Denver’s airport weather sensor—will measure on an hourly basis, the model connects its forecasting task directly to ground-truth observations. This not only provides more granular, locally relevant predictions but also enables better evaluation of the model’s performance against real-world data. Daniel Rothenberg, an atmospheric scientist at Brightband, explained the importance: “Adding a capability where this model is now also predicting, say, what Denver’s airport’s weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core.”
How Does Google Plan to Deploy WeatherNext 3?
This is the first time that core weather variables from a DeepMind model will directly power Google’s consumer and enterprise products. Samier Merchant, a Google senior staff engineer, told TechCrunch that this will be the first time some of the core variables feed into and power a lot of Google products. Users will see the improved forecasts in Search, Google Maps, and the conversational AI Gemini. Developers and researchers will also have access to the model via Google’s cloud platforms. This integration signifies that AI weather forecasting has moved from a research novelty to a practical, production-grade system with global reach.
The Broader Impact: AI Weather Forecasting for the Developing World and Renewable Energy
Beyond product improvements, WeatherNext 3 exemplifies the transformative potential of AI in an area of critical societal importance. Bill Gates recently cited AI-powered weather forecasting as a crucial benefit of the technology, noting that better forecasts can improve crop yields in developing countries. DeepMind’s Ferran Alet, a staff research scientist manager, added that higher-resolution forecasts of wind, rain, and cloud cover will be particularly useful for making renewable energy projects more dependable. Solar and wind farms need accurate short-term forecasts to balance grid loads and optimize power generation. The speed and low cost of AI models also promise to bring economic impact to poorer regions where the expense of high-quality sensors and supercomputers has historically put accurate forecasts out of reach.
The “transformer revolution” that has reshaped language models is now deeply influencing meteorology. European and US weather agencies are already incorporating AI models into their forecast products. As WeatherNext 3 demonstrates, the gap between physics-based and machine learning approaches is closing fast, and in some dimensions, AI has taken the lead. The trend is toward end-to-end models that can learn patterns directly from observations without relying on human-curated physics equations. Alet summarized the core philosophy: “Weather is chaotic, and so small differences really start to perturb massively. Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data.”
WeatherNext 3 represents a clear step toward that ideal. By addressing resolution, rainfall, and temporal frequency, Google has delivered a model that is not just a research achievement but a practical tool ready to shape how billions of people interact with weather information every day. The question is no longer whether AI can compete with traditional forecasting, but how quickly the remaining dependencies on government data can be eliminated, and whether companies like WindBorne and Google can continue to push the boundaries of what is possible with raw observations. The next frontier will be true direct data assimilation, where models learn directly from satellite and sensor streams without intermediate processing. WeatherNext 3 brings us closer to that goal, and the implications for global weather prediction, disaster preparedness, and climate resilience are profound.
