WeatherNext 3 Paper: The Key Points, Explained
WeatherNext 3 was published as a technical report - "WeatherNext 3: Increasing resolution and performance of global weather models with raw observations" (arXiv 2609.03582) - released alongside the September 3, 2026 launch. It is long and dense, so this page distills what it documents, the numbers to look for and the caveats to keep in mind. The paper remains the authority; treat this as a map, not a substitute. For the product view, see the release timeline and the model overview.
7-day free trial, then $19/mo · cancel anytime · compare plans
The numbers the paper turns on
Each figure below is traceable to the paper. The commentary is ours.
1 hour New forecasts every hour instead of every six, by ingesting low-latency geostationary satellite mosaics - the most recent frame arrives with under an hour of latency.
New forecasts every hour instead of every six, by ingesting low-latency geostationary satellite mosaics - the most recent frame arrives with under an hour of latency.
0.1° Single-level variables such as solar radiation and cloud cover resolve at 0.1 degree on hourly time steps, on par with physics-based global models. WeatherNext 2 ran at 0.25 degrees.
Single-level variables such as solar radiation and cloud cover resolve at 0.1 degree on hourly time steps, on par with physics-based global models. WeatherNext 2 ran at 0.25 degrees.
64 Forecasts run as 15-day, 64-member ensembles, with members sampled by injecting noise into the model and skill scored with CRPS across the members.
Forecasts run as 15-day, 64-member ensembles, with members sampled by injecting noise into the model and skill scored with CRPS across the members.
What the paper actually documents
Four threads carry most of the report. Understanding them is most of understanding the paper.
Raw observations in
The model ingests geostationary satellite mosaics - infrared and visible bands covering water vapor, temperature and clouds - plus tropical cyclone fixes and station observations, instead of relying only on analysis fields.
Precipitation from satellite estimates
Beyond classic analysis variables, it learns satellite-derived precipitation products at 0.1 degree, including IMERG-based estimates and a product the team calls PARDIG that its internal checks rank closer to independent observations.
Station-targeted fields
Modelling sparse station data lets 2 m temperature and dewpoint be predicted at any location, conditioned on local geography, with reported error substantially lower than competing global models - including at stations never seen in training.
Evaluation with named baselines
Headline scores use CRPS, the standard probabilistic metric, measured against named references and ensemble baselines such as ECMWF ENS and AIFS ENS, with scorecards comparing model generations at matched resolution.
How to read the paper in about an hour
The report rewards a specific route more than a linear read.
Abstract first, figures second
The abstract names the three headline claims - hourly cadence, on-par resolution, raw-observation training. Go straight to the scorecards before the method.
Read every number with its baseline
Captions say what the model was compared against and at which resolution. A score without its reference tells you little, especially where older forecasts were interpolated for the comparison.
Check the lead times
Skill differences shift with forecast horizon. Look for how far ahead an improvement holds, not just the best row of the best table.
Note the honesty sections
The paper itself flags where evidence is not independent - for example, precipitation verifies partly against products the model also trained on. Read those lines before quoting numbers.
What the paper does not claim
Knowing the edges of a result is as useful as knowing the result.
Not a promise for every town
The reported improvements are strongest where forecasting was historically weakest. Your local baseline may already be strong, so judge the paper by its regional detail, not its headlines.
Not every comparison is like for like
Where generations are compared at 0.1 degree, the older model's forecasts were interpolated upward. That is stated in the paper, but easy to miss in a skim.
Not a safety authority
Nothing in the paper changes who owns warnings: national meteorological agencies remain the source for safety decisions, wherever the forecast comes from.
Not the last word
Evaluation continues after publication, including live leaderboards that rescore model runs as new observations arrive. Treat published numbers as a snapshot.
WeatherNext 3 paper questions
Short answers about the report, its numbers and how to use it.
Related reading: the model overview and accuracy, explained.
From paper to forecast
The paper explains the machine; the workspace shows you its output for your own coordinates. Start the 7-day free trial, or read how to sanity-check accuracy first.
Related guides
The pages that go with this one, chosen by what each one answers.
