Google DeepMind plans to bring WeatherNext 3 forecasts to Search, Maps and Gemini

Google DeepMind and Google Research announced WeatherNext 3, an AI weather forecasting model, on September 3, with plans to distribute its forecasts through Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, Earth Engine, and Google Cloud. The model, led by weather science lead Ferran Alet, claims up to a 60% improvement in Continuous Ranked Probability Score against NASA's IMERG precipitation observations, though independent verification is pending. WeatherNext 3 will be integrated into consumer products and developer services, expanding beyond research use.

Google DeepMind plans to bring WeatherNext 3 forecasts to Search, Maps and Gemini The joint Google DeepMind and Google Research project is moving into consumer products, developer APIs and cloud datasets, where company-reported benchmark gains will meet real weather. By Ryan Merket /author/ryan-merket · Published Primary source: TechCrunch https://techcrunch.com/2026/09/03/googles-latest-ai-weather-model-gives-you-no-excuse-to-forget-your-umbrella/ Why it matters WeatherNext 3 would give a joint Google DeepMind and Google Research forecasting project distribution through Google's consumer products, developer APIs and cloud services. Independent testing will need to show whether Google's reported gains hold during rare, dangerous weather as the planned integrations reach users. Google plans to distribute WeatherNext 3 forecasts through Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, Earth Engine and Google Cloud. Ferran Alet https://alet-etal.com/?ref=runtimewire , Google DeepMind https://deepmind.google/?ref=runtimewire 's weather science lead, and researchers across Google DeepMind and Google Research https://research.google/?ref=runtimewire announced the model https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/?ref=runtimewire on September 3. Alet works in London and earned a computer science PhD from MIT CSAIL after completing degrees in mathematics and engineering physics through Barcelona's UPC-CFIS. Moving WeatherNext 3 into Google's product network would put his weather forecasting research in front of consumers, software developers and companies rather than leaving it inside research papers and benchmark tables. The project sits inside Google DeepMind, the laboratory co-founded and led by Demis Hassabis https://deepmind.google/about/?ref=runtimewire . DeepMind was founded in London in 2010, acquired by Google in 2014 and combined with Google Brain in 2023 to form the current organization. The practical change is distribution: Google intends to place WeatherNext 3 forecasts in Search, Gemini and Maps, while exposing them through the Maps Platform Weather API, Earth Engine and Google Cloud services. Google's previous model provides a technical baseline The public Earth Engine dataset https://developers.google.com/earth-engine/datasets/catalog/projects gcp-public-data-weathernext assets weathernext 3 0 0 0p1deg?ref=runtimewire is listed at 0.1-degree resolution. Google's documentation provides a baseline for the previous generation. WeatherNext 2 https://developers.google.com/weathernext/guides/models?ref=runtimewire uses a 0.25-degree global grid, roughly 30 kilometers at the equator, initializes every six hours and produces 64 ensemble members. The public WeatherNext repository https://github.com/google-deepmind/weathernext?ref=runtimewire documents daily WeatherNext 2 forecast feeds through Google Cloud, including BigQuery. It separately provides pretrained weights and sample data in a Google Cloud Storage bucket. Google says WeatherNext 3 improves Continuous Ranked Probability Score by up to 60% https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/?ref=runtimewire against NASA's IMERG precipitation observations under Google's stated evaluation setup. That figure is not an independent measure of universal forecast accuracy. Brightband's Operational WeatherBench methodology https://owb.brightband.com/methodology?ref=runtimewire lists GraphCast, WeatherNext 2, WeatherNext 3, ECMWF's IFS ENS and AIFS ENS, Microsoft Aurora 1.5, NVIDIA Atlas, GFS, GEFS, HRES and climatology. It says WeatherNext 2 and WeatherNext 3 are initialized from ECMWF IFS analysis and separates deterministic scoring of ensemble means from probabilistic scoring of ensemble members. Fields are evaluated on a fixed 0.25-degree global grid, except WeatherNext 3 surface fields, which are scored at 0.1 degree. Forecasts are evaluated every six hours through 360 hours, or 15 days. Proprietary observations are already part of the competition WindBorne, a weather intelligence company with its own atmospheric sensing network, feeds proprietary measurements from autonomous balloons into WeatherMesh-6 https://windbornesystems.com/blog/introducing-wm-6?ref=runtimewire . In an evaluation covering July 2025 through March 2026 at 0.25-degree resolution, WindBorne says WeatherMesh-6 produced up to 38% lower ensemble-mean root mean square error than ECMWF's physics-based IFS and up to 32% lower error than ECMWF's AIFS https://windbornesystems.com/blog/introducing-wm-6?ref=runtimewire . The company compared WeatherMesh-6's 128-member ensemble with 51-member IFS and AIFS ensembles and defined lower RMSE as higher accuracy. WindBorne said it could not directly compare WeatherMesh-6 with Google's FGN system, making its Google comparison indirect. These are company-reported results and have not been independently verified. Microsoft is also advancing an AI weather model with hourly output. Microsoft released Aurora 1.5 https://www.microsoft.com/en-us/research/blog/aurora-1-5-extending-open-foundation-models-for-weather-and-earth-system-applications/?ref=runtimewire in July 2026 with forecasts covering 22 weather variables and probabilistic ensembles. Comparisons among these systems depend on the selected variables, lead times, observations and error metrics. Alet's research reaches Google's existing products Alet's research page https://alet-etal.com/?ref=runtimewire traces work spanning machine learning for science, GraphCast, GenCast and WeatherNext's cyclone models. RuntimeWire reported in August https://runtimewire.com/article/google-deepmind-open-sources-weathernext-cyclone-models that Google DeepMind's release included WeatherNext 2 and WeatherNext Cyclones. WeatherNext 3 would carry that work into Google's distribution network. Consumers are slated to receive its forecasts through Search, Gemini and Maps, while developers would access them through the Maps Platform Weather API. Google's WeatherNext guidance https://developers.google.com/weathernext?ref=runtimewire directs users to the relevant local meteorological agency or national weather service for official forecasts and warnings. Scientific caution still applies to extreme events. Research on record-breaking weather https://arxiv.org/abs/2508.15724?ref=runtimewire found that physics-based numerical models could outperform leading AI systems on some unprecedented heat, cold and wind events. AI models can smooth rare outcomes or underestimate conditions poorly represented in their training data. Independent evaluations will need to determine whether Google's reported gains persist during rare, dangerous conditions. Once the planned integrations launch across Search, Maps, Gemini, APIs and cloud datasets, forecast errors could reach products and users directly, giving Alet's group a test that research benchmarks alone cannot provide.