WeatherNext 3 Launches on 3 September 2026
On 3 September 2026, Google DeepMind and Google Research launched WeatherNext 3, a global AI weather forecasting model that produces hourly predictions at up to 5-kilometre resolution. Key surface variables — temperature and moisture — are generated at 5km resolution; other surface variables at 10km; and atmospheric variables including wind speed at 25km. This represents a step-change in spatial resolution from the best publicly available numerical weather prediction systems, which typically operate at 25 kilometres for global output. WeatherNext 3 went live simultaneously across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. Researchers and developers can query the full dataset in BigQuery and Earth Engine, or download it in Zarr format from Google Cloud Storage.
The Six-Hour Problem WeatherNext 3 Solves
Traditional numerical weather prediction systems — including ECMWF's ENS, the established global benchmark — update their internal model state every six hours. That update cycle is not a limitation of the physics equations; it is a data collection and assimilation constraint. Radiosondes, aircraft observations, and satellite passes are collected, quality-controlled, and assimilated into the model on a six-hourly schedule. WeatherNext 3 eliminates that constraint by training directly on raw satellite observations rather than on the processed output of traditional models. Satellites stream data continuously; by ingesting each satellite pass as it arrives, WeatherNext 3 generates new forecast updates every hour rather than every six hours.
The architectural shift from physics-based equations to a neural network trained on satellite observations means WeatherNext 3 is not a faster version of the same process. It is a different process — one that learns the patterns of atmospheric change from data rather than encoding them in physical equations. The practical consequence is that the model's predictions reflect real-world atmospheric state far more recently than any traditional NWP output can.
How the Model Handles Uncertainty
WeatherNext 3 uses a diffusion-based generative modelling architecture to produce probabilistic ensemble forecasts rather than single-point predictions. A probabilistic forecast expresses a range of possible outcomes with associated likelihoods, which is substantially more useful for decisions that depend on knowing how confident the forecast is. For infrastructure risk assessment, renewable energy dispatch planning, agricultural water management, or disaster-response pre-positioning, the difference between a deterministic and a probabilistic forecast is often the difference between a useful and an actionable piece of information.
Accuracy Improvements Over Existing Models
WeatherNext 3 outperforms both its predecessor and ECMWF's benchmark system on multiple evaluation metrics. Against weather stations held out from training, the model shows a 30 percent improvement in surface temperature accuracy over WeatherNext 2. At short lead times — forecasts within the first 24 hours — it achieves up to 40 percent better accuracy than ECMWF's ENS for surface temperature. Precipitation forecasting shows the largest gains: WeatherNext 3 improves up to 60 percent against NASA-JPL's IMERG satellite precipitation dataset, 30 percent against MRMS, the US multi-radar multi-sensor analysis, and 10 percent against ground-level rain-gauge measurements at early lead times. For day-ahead and multi-day planning, users will see up to 50 percent more accurate precipitation forecasts, with the largest accuracy gains in regions where traditional NWP models have historically struggled — which includes much of South and Southeast Asia. Independent evaluation by Brightband, a third-party weather model benchmarking service, confirmed these accuracy figures in live production conditions before the public launch.
Data Access for Developers and Researchers
WeatherNext 3 is accessible through four channels. Google Maps Platform's Weather API exposes the forecast data to application developers building location-aware products — logistics platforms, agritech dashboards, event planning services, or outdoor risk management tools. BigQuery provides a SQL-queryable interface for data engineering teams that need to join weather forecast data with structured datasets at scale. Earth Engine provides a geospatial analysis environment for researchers working with large satellite and forecast datasets. For teams that need full global output, Zarr files are available for direct download from Google Cloud Storage, compatible with standard Python scientific computing libraries.
What This Means for Software Teams in India
India is one of the regions where traditional NWP has historically been least reliable. The Indian monsoon is a complex atmospheric system that global models built primarily on data from Europe and North America represent imperfectly, especially at the sub-50-kilometre scale needed for practical agricultural, logistics, and disaster-response decisions. A 5-kilometre resolution model that updates every hour from live satellite data is materially more useful in an Indian context than the best previously available managed APIs.
For teams building agritech, supply chain, insurance, or disaster-response products, WeatherNext 3's availability via the Google Maps Platform Weather API and BigQuery changes the underlying data quality of their product without requiring a custom satellite data pipeline. Teams that have previously relied on India Meteorological Department data or third-party weather APIs can now access a model that demonstrably outperforms existing sources on the metric — precipitation accuracy — that matters most for Indian agricultural and flood-prediction applications. The BigQuery interface is particularly relevant for Indian data engineering teams: rather than managing raw satellite data ingestion pipelines, teams can query global hourly forecast data at 5-kilometre resolution directly in SQL, joining it with their own datasets without the operational overhead of a weather data pipeline.
The Bottom Line
Google DeepMind launched WeatherNext 3 on 3 September 2026. The model generates global hourly weather forecasts at up to 5-kilometre resolution for key surface variables, training directly on live satellite observations to update every hour rather than on the six-hourly cycle of traditional numerical weather prediction. Accuracy improvements include a 30 percent gain in surface temperature over WeatherNext 2, up to 40 percent over ECMWF's ENS at short lead times, and up to 60 percent better precipitation accuracy versus NASA-JPL's IMERG. The data is integrated into Google Search, Gemini, Google Maps, and the Maps Platform Weather API, and is accessible to developers via BigQuery, Earth Engine, and Zarr download from Google Cloud Storage. For software teams in India building weather-dependent products, the combination of 5km resolution, hourly updates, and demonstrated precision gains in precipitation makes WeatherNext 3 the most practically useful managed weather data source now available.
Frequently Asked Questions
What is WeatherNext 3 and when was it released?+
WeatherNext 3 is a global AI weather forecasting model released by Google DeepMind and Google Research on 3 September 2026. It generates hourly weather predictions at up to 5-kilometre resolution for key surface variables including temperature and moisture, 10km for other surface variables, and 25km for atmospheric variables such as wind speed. Unlike traditional numerical weather prediction systems, WeatherNext 3 trains directly on raw satellite observations and updates every hour rather than on the six-hourly cycle that has constrained weather forecasting for decades. The model uses a diffusion-based generative architecture to produce probabilistic ensemble forecasts, expressing uncertainty as a range of likely outcomes rather than a single point prediction.
How does WeatherNext 3 compare to ECMWF and traditional weather models in accuracy?+
WeatherNext 3 outperforms both its predecessor WeatherNext 2 and ECMWF's ENS benchmark across multiple metrics. Against held-out weather stations, it achieves a 30 percent improvement in surface temperature accuracy over WeatherNext 2 and up to 40 percent better accuracy than ECMWF's ENS at short lead times. Precipitation accuracy — the variable most critical for agriculture, flood prediction, and logistics — improves by up to 60 percent versus NASA-JPL's IMERG satellite precipitation dataset, 30 percent versus MRMS, and 10 percent versus rain-gauge measurements at early lead times. For day-ahead and multi-day forecasts, users see up to 50 percent better precipitation accuracy in products integrated with WeatherNext 3. Accuracy gains are largest in regions where traditional models have historically struggled, including South and Southeast Asia. Independent evaluation by Brightband confirmed these figures in live production before launch.
How can developers and data engineers access WeatherNext 3 data?+
WeatherNext 3 data is accessible through four channels as of its 3 September 2026 launch. The Google Maps Platform Weather API provides a managed endpoint for application developers building location-aware products such as logistics, agritech, or risk management services — no data pipeline management required. BigQuery provides a SQL-queryable interface for data engineering teams who need to join hourly 5km forecast data with their own structured datasets at scale. Google Earth Engine offers a geospatial analysis environment for researchers working with large satellite and climate datasets. Finally, full global hourly forecast output is available for direct download in Zarr format from Google Cloud Storage, compatible with standard Python scientific computing libraries such as xarray.
What does WeatherNext 3 mean specifically for software teams and businesses in India?+
India is one of the regions where traditional numerical weather prediction has historically been least accurate, particularly for monsoon precipitation at the sub-50-kilometre scale needed for practical agricultural and flood-response decisions. WeatherNext 3's 5km resolution and hourly satellite-driven updates provide a materially better data source than any previously available managed weather API for Indian use cases. For teams building agritech platforms, crop insurance products, supply chain logistics systems, or disaster-response tools, WeatherNext 3 data via the Google Maps Platform Weather API or BigQuery is now the highest-accuracy managed source for precipitation — the most critical weather variable for Indian applications. Teams that previously relied on India Meteorological Department data or third-party weather APIs can access WeatherNext 3 without building custom satellite data pipelines.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
