Google WeatherNext 3: Hourly AI Forecasts, 5-Kilometer Resolution, Gemini Integration, and Real-Time Satellite Data

Google DeepMind and Google Research launched WeatherNext 3 on September 3, 2026 as a new global AI weather model built around faster refresh cycles, higher spatial resolution, real-time satellite observations, station-level targets, and direct deployment across Google products and cloud data services.

The system is designed to reduce two practical weaknesses of earlier AI forecasting pipelines: coarse local detail and the delay created when models depend mainly on numerical weather prediction analyses that are produced on slower update cycles. WeatherNext 3 instead ingests live geostationary satellite mosaics alongside historical analysis and can issue a new forecast every hour.

Google says the model produces a global weather picture roughly five times sharper than WeatherNext 2, while independent live evaluations from Brightband place it among the strongest systems tested on operational metrics. Those claims need to be read variable by variable, because resolution, precipitation skill, atmospheric fields, and official warning responsibilities are separate questions.

··········

WEATHERNEXT 3 MOVES FROM 25-KILOMETER, SIX-HOUR OUTPUTS TO VARIABLE RESOLUTION AND HOURLY FORECASTS.

The largest product-level change is the combination of finer spatial detail with a much faster refresh cadence.

WeatherNext 2 generated forecasts on a 25-kilometer grid in six-hour increments. WeatherNext 3 can represent key surface variables such as temperature and moisture at 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers, while issuing updated forecasts every hour.

That distinction matters operationally because the headline 5-kilometer figure does not apply uniformly to every output field. The model uses multiple resolutions to preserve global-scale consistency while adding more local detail where the architecture and training targets support it.

........

Capability	WeatherNext 2	WeatherNext 3
Forecast refresh	Every 6 hours	Every hour
Key surface variables	25 km grid	Up to 5 km
Other surface variables	25 km grid	10 km
Selected atmospheric variables	25 km grid	25 km
Live satellite observations	Not the central real-time input path	Hourly geostationary satellite mosaics are ingested
Station-level targets	More limited	Native forecasts at sparse station coordinates

........

The five-times-sharper framing is therefore a broad system-level comparison rather than a statement that every forecast variable is always produced at 5 kilometers. For users evaluating the model, the relevant resolution is the one attached to the specific variable and product surface they intend to use.

··········

REAL-TIME SATELLITE DATA CHANGES THE INPUT PIPELINE, BUT IT DOES NOT REMOVE TRADITIONAL WEATHER DATA FROM THE SYSTEM.

WeatherNext 3 brings raw observations closer to the forecasting loop, while still operating inside a hybrid data ecosystem.

Most AI weather systems are trained heavily on reanalysis or numerical weather prediction datasets generated by physics-based models. Google says those pipelines can introduce a data lag of roughly six hours, which is particularly costly for fast-changing variables such as precipitation and surface temperature.

WeatherNext 3 ingests one-hour global geostationary satellite mosaics and combines them with traditional historical analysis inside a Functional Generative Network mesh-transformer architecture. The model then produces dense gridded fields, cyclone tracks, and station-level sparse outputs from the same system.

Google also trains parts of the system directly against sparse weather-station observations. That gives the decoder targets tied to specific ground locations, which can improve local usefulness and make evaluation against observed station measurements more direct.

The precipitation pipeline uses NASA IMERG satellite precipitation data and Google's own global precipitation reanalysis. Google reports medium-range CRPS improvements of up to 60% against IMERG, 30% against MRMS, and 10% against rain-gauge measurements for early lead times. These are model evaluation results reported by Google and should be treated as benchmark-specific performance claims rather than universal percentage gains for every location or event.

TechCrunch also notes an important qualification around Google's description of WeatherNext 3 as the first AI model to directly incorporate raw observations for a high-resolution global forecast. WindBorne disputes the broader first-mover framing because its WeatherMesh 6 system has incorporated raw observations since late 2025. Both approaches still use national weather datasets elsewhere in the pipeline, so direct observation ingestion does not yet mean a completely independent end-to-end data-assimilation stack.

··········

GOOGLE IS DEPLOYING THE MODEL ACROSS SEARCH, GEMINI, MAPS, CLOUD, AND CLEAN-ENERGY WORKFLOWS.

WeatherNext 3 is being treated as production infrastructure rather than a research model that remains isolated from user-facing products.

Starting on launch day, Google said WeatherNext 3 would begin powering weather experiences in Google Search, the Gemini app, Google Maps, Google Maps Platform Weather API, and Google Earth Engine. High-resolution forecasts are also available through BigQuery, Earth Engine, and bulk download from Google Cloud Storage for developers, researchers, and businesses.

The model also adds variables designed for renewable-energy operations, including 100-meter wind speeds near turbine height, cloud cover, and surface solar radiation. These outputs can be used to estimate wind and solar generation and to improve short-term balancing decisions where weather uncertainty directly affects grid planning.

........

Deployment surface	What WeatherNext 3 adds	Practical implication
Google Search	Higher-resolution and updated forecasts	Consumer weather queries can use the new model directly
Gemini app	Weather intelligence from WeatherNext 3	Conversational planning can use fresher forecast data
Google Maps	Improved weather experiences	Route and destination context can include better forecast information
Google Maps Platform Weather API	Hourly global forecast data	Developers can integrate the model into applications
BigQuery / Earth Engine / Cloud Storage	Queryable and downloadable forecast datasets	Research and enterprise workflows can consume model output without running the model
Renewable energy	100 m wind, cloud cover, solar radiation	Generation forecasting can be tied more closely to operational assets

........

Google says longer-term precipitation forecasts shown to users can be up to 50% more accurate, with the largest gains in regions where prior forecasts were less reliable. That figure is a Google product-performance claim and should not be interpreted as a uniform improvement across all geographies, lead times, variables, or hazardous-weather scenarios.

··········

WEATHERNEXT 3 IS MOST CONSEQUENTIAL WHERE FRESHER OBSERVATIONS AND LOCAL DETAIL CHANGE A REAL DECISION.

The technical advantage is strongest when the value of a forecast depends on both update frequency and spatial specificity.

For agriculture, logistics, aviation, emergency response, renewable energy, and consumer planning, a forecast that refreshes every hour can react to newly observed atmospheric structure faster than a pipeline centered on six-hour analysis cycles. The higher-resolution surface outputs can also reduce the smoothing that makes coastlines, valleys, mountains, and urban temperature patterns difficult to represent on coarser grids.

The model still does not replace national meteorological agencies for official warnings or safety-critical decisions. Google explicitly directs users to local meteorological agencies and national weather services for severe-weather alerts and public-safety advisories, and the atmosphere remains intrinsically uncertain even when model skill improves.

The practical benchmark for WeatherNext 3 is therefore not whether one headline score beats every traditional forecast, but whether its hourly observation cycle, variable-specific resolution, precipitation improvements, and production integrations create better decisions in the exact workflow where the forecast is consumed. Its strongest contribution is the convergence of AI forecasting research with a global delivery stack that can expose those improvements directly to consumers, developers, and operational systems.

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]