WeatherNext 3 vs GraphCast: Three Generations Compared
GraphCast proved a learned model could match the physics; GenCast turned that into ensembles; WeatherNext 3 turned it into an hourly feed you can use before a commute. This comparison lines the three up on the dimensions that decide real use - cadence, resolution, ensembles and inputs - and closes with which job each one still fits. The model's own numbers live in the paper walkthrough and the accuracy breakdown.
7-day free trial, then $19/mo · cancel anytime · compare plans
The short version
One DeepMind project in three acts, each solving the problem the last one left open.
GraphCast, 2023 The precision milestone: a learned model that matched the best physics forecasts on most variables. Deterministic, 0.25 degree, refreshed every six hours.
The precision milestone: a learned model that matched the best physics forecasts on most variables. Deterministic, 0.25 degree, refreshed every six hours.
GenCast, 2024 The ensemble turn: a diffusion model that produced a 50-member ensemble and turned forecasts into distributions rather than single numbers.
The ensemble turn: a diffusion model that produced a 50-member ensemble and turned forecasts into distributions rather than single numbers.
WeatherNext 3, now The operational generation: hourly updates, 0.1 degree precipitation, 64-member ensembles and live satellite inputs - built to be used, not just benchmarked.
The operational generation: hourly updates, 0.1 degree precipitation, 64-member ensembles and live satellite inputs - built to be used, not just benchmarked.
The comparison that actually matters
Four dimensions explain almost all of the difference in daily use.
Cadence: six-hourly against hourly
GraphCast and GenCast refresh every six hours. WeatherNext 3 is designed to regenerate every hour, which is what makes within-day planning - a commute, a shift, a launch window - practical.
Resolution: 0.25 against 0.1 degrees
The earlier models work on a 0.25 degree grid; WeatherNext 3 resolves precipitation at 0.1 degree, with targeted temperature and humidity at 5 km and other surface variables at 10 km.
Ensembles: 50 members against 64
GenCast popularised learned ensembles with 50 members; WeatherNext 3 runs 15-day, 64-member ensembles, sampled so the spread stays honest.
Inputs: reanalysis against live satellite
GraphCast and GenCast learn from historical reanalysis. WeatherNext 3 draws directly from raw satellite imagery, which is what expands its reach and detail.
Picking a model in four questions
A short decision routine if you are choosing where to put attention or compute.
Is the decision inside a day?
If yes, cadence decides first: hourly updates beat a finer grid when the plan keeps changing during the day.
Do you need to see uncertainty?
GraphCast alone is deterministic. GenCast and WeatherNext 3 return ensembles, which is the difference between one number and a range of futures.
Do you want the weights?
GraphCast is famously open, and the WeatherNext family publishes weights for earlier generations plus hosted data - the GitHub guide lists what is available.
Run it, or use it?
Running a model takes compute and pipelines; using a forecast takes a browser. This site is the second path, which is the point of the product.
Same lineage, different jobs
Each generation still earns its place; none of them retired the last one.
GraphCast walked so ensembles could run
The 2023 result is why the field started trusting learned models with operational questions.
GenCast made uncertainty first-class
Turning forecasts into distributions changed what people could ask of a model at all.
WeatherNext 3 made it a product
Hourly cadence, sharper fields and satellite inputs are the step from research result to something you check before leaving home.
The older models still matter
Open weights and years of tooling keep GraphCast and GenCast as research baselines; WeatherNext 3 is where the hour-by-hour experience lives.
GraphCast and GenCast questions
The comparisons people actually search for, answered without hedging.
Next: what changed in the model or the GitHub guide.
Two generations of difference, one click away
Read what changed under the hood, then judge the hourly view on your own weather. Seven-day free trial, cancel anytime. Background: the paper walkthrough.
Related guides
The pages that go with this one, chosen by what each one answers.
WeatherNext 3 paper
The key points of arXiv 2609.03582: hourly satellite-driven forecasts, 0.1 degree fields and 64-member ensembles.
WeatherNext 3 model
Architecture, resolution and what changed between WeatherNext generations.
WeatherNext 3 accuracy
What the published evaluations show, and how to read them.