WeatherNext 3: DeepMind's New AI Weather Model Goes Hourly at 5km Resolution — What Engineers Need to Know
DeepMind dropped WeatherNext 3 on September 3, 2026, and it's the first global AI weather model that generates forecasts every single hour. The Hacker News reception (236 points) is deserved — this isn't an incremental update. It's an architectural change in how AI weather models ingest data, and it has real implications for renewable energy, aviation, disaster response, and every Google product you use to check the sky.
Here's how it works, where it beats numerical weather prediction, where it doesn't, and why the German Meteorological Service thinks the future is hybrid.
What WeatherNext 3 Actually Is
WeatherNext 3 is an ensemble AI model that predicts global weather by learning directly from observations rather than relying on physics equations or pre-processed analysis fields. It's the successor to WeatherNext (2024) and WeatherNext 2 (2025), and it changes two fundamental things about how the model works:
Direct satellite ingestion. Previous WeatherNext runs used ERA5 reanalysis data — the ECMWF's best-guess historical reconstruction of the atmosphere. WeatherNext 3 draws directly from raw satellite imagery (infrared, visible, water vapor channels) and station observations. That means it sees what the satellite actually sees, not what a physics model says the satellite should have seen.
Hourly forecast cadence. Global NWP models typically produce forecasts every 3 or 6 hours (ECMWF HRES at 6h, GFS at 3h). WeatherNext 3 updates every hour. For tracking fast-moving phenomena — convective storms, squall lines, rapid cyclogenesis — that's the difference between knowing a front passed through at 2pm vs "sometime this afternoon."
The Resolution Story
| Variable | Resolution | What Changed |
|---|---|---|
| Temperature, humidity (at station height) | 5 km | Previously 10–20 km in AI models |
| Wind, pressure, other surface variables | 10 km | Comparable to ECMWF HRES |
| Forecast frequency | Every 1 hour | Previously 3–6 hours (NWP) or 6–12 hours (AI) |
| Update cadence | Multiple runs per day | Driven by real-time satellite feeds |
The 5 km temperature/humidity resolution is the headline number here. Most global AI weather models (GraphCast, Pangu-Weather, FourCastNet) operate at roughly 0.25° — about 28 km at the equator. WeatherNext 3's 5 km (roughly 0.045°) is a 5–6x linear resolution improvement over previous-gen AI models in the variables that matter for local forecasting.
The DWD (German Meteorological Service) also notes that directly using radar and satellite observations means the initial precipitation field is much closer to reality than an NWP analysis, which has to blend sparse observations with a background model. This gives WeatherNext 3 a structural advantage on precipitation nowcasting.
How It Works (Without the Full Paper)
The research paper (22 MB PDF on Google Storage) describes an architecture that builds on DeepMind's GraphCast lineage but with significant changes to the input encoding and output head:
- Graph neural network backbone with learned mesh representation at multiple scales
- Direct observation encoder that maps satellite radiance channels and station METAR data into the latent space, bypassing the data-assimilation pipeline that traditional NWP requires
- Ensemble output — multiple forecast members quantify uncertainty, which is critical for probabilistic decision-making in energy trading, aviation, and emergency management
- Temporal conditioning that allows the model to ingest the latest observation at any hour and produce a forward rollout without waiting for a global analysis cycle
Where It Wins: Real Evidence
Hurricane Melissa. DeepMind's announcement notes that WeatherNext 3 was used by the US National Hurricane Center to predict Hurricane Melissa's historic landfall in Jamaica. That's an operational validation by the agency responsible for tropical cyclone forecasting in the Atlantic — not a self-reported benchmark.
Renewable energy. The model explicitly forecasts radiation and cloud cover variables designed for wind and solar farm operators. For anyone trading energy based on weather forecasts, the hourly update cycle and direct satellite cloud cover estimates are materially better than the 3–6 hour lag of NWP alternatives.
Precipitation nowcasting. Because the model starts from raw radar/satellite data instead of an analyzed field, the first few hours of the precipitation forecast are structurally better than NWP. Radar sees actual rain. NWP has to infer it from modeled physics.
Where It Falls Short (DWD's Independent Analysis)
The German Meteorological Service published an unusually detailed evaluation. Here are the weaknesses they identified:
1. Measurement errors aren't explicitly modeled. NWP data assimilation systems explicitly model the error covariance of every observation type. A neural network has to learn these error structures implicitly from training data. If a sensor drifts, fails, or gets replaced with a new instrument, the model has no mechanism to know the error budget changed.
2. Climate change creates unseen regimes. Neural models learn from historical data. If the atmosphere enters a state that doesn't exist in the training distribution — higher baseline temperatures, different moisture patterns — the model can systematically underestimate extremes. An NWP model captures this through the Clausius-Clapeyron relation. An AI model has to have seen it in training.
3. Physical consistency isn't guaranteed. NWP models solve coupled equations — mass, momentum, energy are conserved by construction. An AI model can produce outputs where each variable looks plausible individually but the combination violates basic conservation laws.
4. Spatial artifacts from the neural architecture. The model's internal grid and graph structure can systematically prefer certain spatial scales or symmetries, creating smoothing or distortion — especially in small-scale features like convective cells and frontal boundaries.
5. Interpretability is worse. When an NWP forecast goes wrong, you can trace the error to a specific parameterization, initial condition, or numerical instability. With a neural model, diagnosing a bad forecast requires understanding which training examples drove the behavior.
The Integration You'll Actually Use
WeatherNext 3 is already being rolled into:
- Google Search — your "weather near me" query now runs on WeatherNext 3 outputs
- Google Maps — route planning with hourly precipitation forecasts
- Gemini — "will it rain at 3pm?" queries
- BigQuery and Earth Engine — enterprise API access for developers who need to query forecasts programmatically
- Google Maps Platform — fleet routing, logistics optimization
- Weather Lab (weatherlab.deepmind.google) — experimental interactive viewer for comparing model outputs
The Hybrid Future
The DWD's conclusion is the most honest framing I've seen: AI weather models and numerical weather prediction are complementary, not competing:
| Aspect | NWP (ECMWF, GFS, ICON) | AI (WeatherNext 3) |
|---|---|---|
| Physics conservation | Built-in | Learned, not guaranteed |
| Observation usage | Via data assimilation (hours to run) | Direct ingestion (minutes) |
| Forecast speed | Hours on supercomputers | Minutes on GPUs |
| Extreme events in novel climate | Handled by physical laws | Only if in training distribution |
| Interpretability | Traceable to parameterization | Black box |
| Precipitation nowcasting | Good after spin-up | Excellent from hour 0 |
| Resolution | ~9 km global (ECMWF) | 5–10 km global |
The likely outcome: hybrid systems where AI models provide the high-frequency, high-resolution short-range forecast, while NWP models provide the physically consistent boundary conditions and handle the long-range and climate-extrapolation cases.
What Engineers Should Do
If you're building on weather data:
- For solar/wind energy operations: Test WeatherNext 3's cloud cover and radiation forecasts against your current NWP provider. The hourly update cycle alone may justify switching.
- For logistics routing: The Maps Platform integration means you can skip the weather API integration and use Google's own routing with WeatherNext 3 baked in.
- For aviation or emergency management: Pair AI forecasts with NWP ensembles for high-stakes decisions. Neither alone is sufficient.
- For research: The Weather Lab explorer lets you compare WeatherNext 3 against ECMWF and GFS baselines interactively.
The paper, the interactive lab, and the enterprise APIs are all live today. WeatherNext 3 is not a research preview — it's in production in Search, Maps, and Gemini right now. For short-range, high-resolution forecasting, it's the strongest AI weather model anyone has shipped.
Sources: DeepMind WeatherNext 3 announcement, Research paper (PDF), DWD independent evaluation (German), HN discussion (236 pts), Weather Lab interactive explorer.
Top comments (0)