DeepMind launches WeatherNext 3 with hourly, 5km AI forecasts

DeepMind's new weather model refreshes its forecast every hour at 5-kilometre resolution, a fivefold jump in spatial detail and a sixfold jump in cadence over the previous release. WeatherNext 3, built with Google Research, pulls live geostationary satellite imagery on each cycle and trains on ground-station observations to keep its view of the atmosphere honest about what is actually happening at the surface.

The headline numbers come from precipitation. Google reports up to 50% better day-ahead rain and snow forecasts than WeatherNext 2, with certain precipitation benchmarks improving by as much as 60%. The previous model ran on a 25-kilometre grid and refreshed every six hours; hourly updates at 5km matter most where storms form and decay on the timescale of a coffee break, which is exactly the regime where the old cadence was useless.

The model is already wired into the consumer stack. WeatherNext 3 now feeds the Gemini app, Google Search, Google Maps, Google Maps Platform, and Google Cloud. Beyond weather planning, Google says the same pipeline can drive agricultural decisions and renewable-energy scheduling by forecasting wind speeds, cloud cover, and solar radiation at the same temporal and spatial granularity.

The practical angle

For practitioners, the meaningful question is whether a model trained on live satellite radiances and surface stations still degrades gracefully a few days out. AI weather systems built purely on reanalysis have a well-known horizon problem: skill falls off a cliff past about five to seven days, often worse than the ECMWF IFS for medium-range synoptic features. Tightening the grid and the refresh interval buys you nowcasting, which is exactly the slice where operational forecasting still loses the most money, but it does not by itself fix medium-range bias.

Independent evaluation matters more than vendor numbers here. TechCrunch has flagged that WeatherNext 3 has held up in third-party comparisons, and that is a signal worth tracking as the model rolls out. Anyone integrating these forecasts into a downstream pipeline, from grid operators to ride-share dispatch, should still run a shadow mode against their current numerical baseline for at least one season before swapping anything in.

Where this fits in the AI weather race

WeatherNext 3 sits inside a broader migration from physics-based numerical weather prediction to hybrid and data-driven systems, a shift that NVIDIA's Earth-2, Microsoft Aurora, and the various ECMWF AI pilots are all chasing. The differentiator DeepMind is selling is operational density: not a bigger model, but a shorter loop between observation and forecast, served through APIs that an application team can hit directly. That is also why Google routes the model through Cloud and Maps rather than only through a research portal. Distribution is the moat, not the architecture.

For ML engineers watching the space, two engineering details are worth filing away. First, ingesting geostationary data on every cycle rather than caching a fixed training corpus means the system has to handle streaming multimodal inputs at production scale. Second, training on station observations alongside satellite fields is a quiet admission that pure reanalysis-only models underfit near-surface variables, which is the same correction Aurora and several recent diffusion-based weather models have made.

It is still early days for this generation of AI weather models, and vendor benchmarks remain the loudest signal in the room. The honest read is that WeatherNext 3 looks strong on short-range, high-impact precipitation, where the resolution and cadence line up with the physics, and unproven on the medium-range tail, where the community will need months of head-to-head data against IFS and the GFS before the case closes.

Frequently asked questions

What is WeatherNext 3?
It is the latest AI weather model from Google DeepMind and Google Research, generating global forecasts at 5km resolution on an hourly cadence using live geostationary satellite data and ground station observations.

How is it different from WeatherNext 2?
WeatherNext 2 ran at 25km resolution and refreshed every six hours. WeatherNext 3 is five times finer in space and six times more frequent in time, with reported precipitation accuracy gains of up to 50 to 60%.

Where can I access it?
Forecasts are integrated into Google Search, the Gemini app, Google Maps, Google Maps Platform, and Google Cloud. API access for developers goes through Google Cloud.

Is WeatherNext 3 better than traditional numerical weather models?
For short-range, high-resolution precipitation and nowcasting it appears competitive and in some cases ahead, but the medium-range picture versus systems like ECMWF IFS is still being evaluated and is not settled.