AI weather models have spent three years closing the gap with physics-based forecasting, but two problems stayed open: resolution too coarse for local terrain, and initialization tied to numerical weather prediction (NWP) analysis that arrives about six hours late. WeatherNext 3, released by Google DeepMind and Google Research, attacks both. It takes a live global geostationary satellite mosaic as a direct model input, re-initializes every hour, and emits forecasts down to 0.05° (~5 km) while training against raw weather station measurements rather than reanalysis grids alone. Independent live evaluations from Brightband rank it as the most accurate global weather model to date, according to Google AI. Forecast data is available now through BigQuery, Earth Engine and Cloud Storage after an allowlist request, but WeatherNext 3 weights are not open source and on-demand custom inference still runs WeatherNext 2.
Architecture and Inputs: A Functional Generative Network That Ingests Satellite Data in Real Time
WeatherNext 3 is a Functional Generative Network (FGN) mesh transformer, the same probabilistic family introduced with WeatherNext 2, scaled to multi-resolution output. Inputs are a live geostationary satellite mosaic plus ECMWF HRES analysis. Training draws on ERA5/HRES-fc0, NASA’s IMERG, station observations and satellite mosaics. Most AI forecasters learn from NWP reanalysis, which smooths away the local variation that coastlines, valleys and mountains actually produce. WeatherNext 3 trains dedicated observational heads directly on raw station measurements, so its 0.05° temperature and dew point outputs are calibrated to what instruments record rather than to a model’s representation of the atmosphere.
The shift from a six-hourly NWP-anchored initialisation to a 24-times-daily satellite-fed cycle is the most consequential architectural change. Numerical weather prediction analysis itself introduces latency: the data must be collected, assimilated, and run through a physics-based system before it can seed an AI model. By pulling a live geostationary mosaic directly into the transformer, WeatherNext 3 essentially bypasses that queue. The model can now capture rapidly developing weather events — such as convective storms, sea-breeze fronts, or fog formation — that would otherwise be smoothed over in a six-hour window.
Multi-Resolution Output: From 0.05° Station Variables to 0.25° Pressure Levels
A single forward pass produces three tiers. The first tier is 0.05° (~5 km) station-trained 2 m temperature and dew point, calibrated directly to instrument readings. The second tier is 0.1° (~10 km) gridded surface variables: 10 m and 100 m wind, pressure, sea surface temperature, cloud layers, solar radiation, and 1-hour precipitation. The third tier is 0.25° (~25 km) atmospheric fields across 13 pressure levels. WeatherNext 2 produced 0.25° fields in 6-hour increments, which is where the roughly 5x sharper claim comes from.
This multi-resolution strategy is pragmatic. Full-atmosphere 0.05° output at every time step would be computationally prohibitive for a global model running 24 times a day. Concentrating the fine grid on surface variables that directly affect human activity — temperature, humidity, wind — while keeping the synoptic-scale fields at a coarser resolution allows the model to allocate compute where it matters most. For a grid operator trying to forecast afternoon temperature spikes or a farmer watching dew point thresholds, the 0.05° station-trained output is a meaningful step beyond the 25 km block averages that dominated AI weather modelling until now.
Hourly Cadence and Ensemble Design: A 64-Member Forecast That Refreshes Every 60 Minutes
Cadence is the second change. The model initializes 24 times a day. The 00, 06, 12 and 18 UTC synoptic cycles run out to 15 days (360 hours) with 64 ensemble members; interim hourly runs cover 48 hours. For fast-developing convection, an hourly refresh grounded in current satellite observations is meaningfully different from a six-hourly cycle anchored to lagged analysis. The 64-member ensemble provides probabilistic spread that captures forecast uncertainty at longer lead times, while the short-range hourly runs can be treated as deterministic for near-term decisions.
Questions from operational users typically focus on how this ensemble compares to established ones. What is the WeatherNext 3 ensemble size? 64 members for the synoptic cycles. How often does it update? Every hour for the first 48 hours, and every six hours out to 15 days. The 48-hour interim runs are designed for applications that need fresh nowcast-like information — aviation dispatch, construction scheduling, event planning — where a 24-hour-old model initialisation is already outdated.
Precipitation Forecasting: Where WeatherNext 3 Claims the Largest Gains
Precipitation is where global models historically fail, producing blurred fields that miss storm boundaries. WeatherNext 3 trains against three precipitation sources: ECMWF reanalysis, NASA’s IMERG satellite retrievals, and Google’s own satellite-radar precipitation reanalysis. The combination allows the model to learn the sharp edges of convective bands rather than smearing them into diffuse blobs.
Google reports CRPS improvements over baselines of up to 60% against IMERG, 30% against MRMS and 10% against rain gauges at early lead times; the research separately states up to a 50% reduction in Brier score and CRPS versus NWP baselines when evaluated against IMERG. These numbers are notable because IMERG itself is a satellite-based product with its own uncertainties; beating it by 60% in the first few hours suggests that the model is capturing sub-grid-scale structure that even high-resolution satellite retrievals can miss. The 50% improvement over physics-based NWP at day-plus lead times is equally significant for medium-range planning in sectors like agriculture and energy.
What does CRPS mean in this context? The Continuous Ranked Probability Score measures the quality of probabilistic forecasts — lower is better. A 60% reduction means the forecast probability distribution is much closer to the observed outcome. For a utility company deciding whether to pre-position repair crews, that translates directly into fewer false alarms and fewer missed events.
Clean Energy Variables: A Clear Signal of Operational Intent
For renewables, the model outputs 100 m wind speed at approximate turbine hub height, full low/medium/high cloud distributions, and both solar irradiance components (SSRD and FDIR). That combination is what grid operators need to forecast wind and solar output against demand, and it is the clearest sign that this release is aimed at operational buyers, not only at benchmark tables. Temperature and dew point at 2 m, along with surface pressure, complete the set of variables most critical for load forecasting and thermal plant efficiency.
The 100 m wind output is especially relevant because many AI weather models still produce 10 m wind, which is not directly usable for turbine power curves. By offering an explicit hub-height variable, WeatherNext 3 removes a step that operators previously had to handle with boundary-layer scaling formulas. Cloud distributions at three levels allow solar irradiance models to account for low, middle, and high cloud separately — a level of detail that most NWP models provide but that AI models have only recently begun to match.
Deployment and Access: Available by Request, Not Open Weights
WeatherNext 3 forecast data is available now through BigQuery, Earth Engine and Cloud Storage after an allowlist request. The model itself is not open source, and on-demand custom inference still routes through WeatherNext 2. That dual track means researchers and commercial users can access the output without needing to run the model themselves, but the inability to fine-tune or modify the weights limits adoption in specialised applications. Google’s approach mirrors that of other major AI weather efforts: the inference infrastructure is cloud-native and gated, while the weights remain proprietary.
Brightband, an independent evaluation platform, ranks WeatherNext 3 as the most accurate global weather model to date based on live runs. That ranking covers a range of variables and regions, though the exact methodology and weighting are not fully public. For organisations that can navigate the allowlist process, the data is immediately usable through Google Cloud tools — BigQuery for SQL-based analysis, Earth Engine for geospatial visualisation, and Cloud Storage for bulk download.
How WeatherNext 3 Compares to Physics-Based Models and Its Own Predecessor
The direct comparison with ECMWF’s HRES, the gold standard for global NWP, is complicated by the fact that WeatherNext 3 uses HRES as one of its inputs. A better baseline is the model’s own predecessor. WeatherNext 2 operated at 0.25° with six-hour time steps and no satellite ingestion. The jump to 0.05° for surface variables, hourly updates, and direct observational training is a generational leap, not an incremental improvement. The 64-member ensemble also represents a substantial increase from the smaller ensemble sizes typical of earlier AI models, which often ran deterministic forecasts or small perturbed members.
What makes WeatherNext 3 genuinely deployable today? The data pipeline is already integrated into Google Cloud, and the output variables align with real-world decision-making: temperature, dew point, wind at turbine height, solar irradiance, and precipitation with sharp convective edges. The hourly refresh cadence means that users are never more than 60 minutes away from a new forecast, which is essential for operational weather-sensitive industries. The 15-day horizon on the synoptic cycles, while less frequent, provides the medium-range outlook that commodity traders and logistics planners rely on.
The primary limitation is the closed-weight status. Open-source alternatives like FourCastNet, GraphCast, and Pangu-Weather have published weights, enabling custom inference, fine-tuning, and local deployment. WeatherNext 3’s data-as-a-service model may appeal to organisations that prefer not to manage model infrastructure, but it will not replace on-premise forecasting for air-gapped or latency-sensitive applications. For those users, WeatherNext 2’s on-demand inference remains the available option.
Looking ahead, the integration of live satellite imagery as a direct model input is likely to become standard across the field. The next frontier is moving beyond geostationary mosaics to assimilating radar, lightning, and aircraft observations in real time. WeatherNext 3 demonstrates that the architecture is ready for that transition. The model’s performance on precipitation and the explicit clean-energy variables also suggest that the business case is clear: energy and agriculture are the verticals that will drive adoption, not just meteorology. As cloud-based AI weather models mature, the question will shift from “can they beat NWP?” to “can they meet the data rights, latency, and customisation requirements of a thousand different operational contexts?”. WeatherNext 3 answers the first question affirmatively, while leaving the second open for the next release.
