14 models, and what separates them

Architecture

Graph transformer, encoder–processor–decoder

Grid

0.25° global

Horizon

360 h

Cadence

00 / 06 / 12 / 18Z

Access

Open data from ECMWF, mirrored on AWS, Azure and Google Cloud; GRIB2 with a JSON index sidecar

Licence

CC-BY-4.0

It is the control. Same centre, same open-data plumbing, same cycles as the IFS it sits beside — so AIFS minus IFS is a genuinely clean subtraction, with no difference in resolution, format or timing to explain away.

The first learned model any national centre ran operationally, and still the easiest to build on: the open-data layout is byte-range friendly, so a map costs a few MB rather than the whole file. Our engine already reads it — weathermaps/ecmwf_opendata.py handles ifs and aifs-single through one path.

Architecture

Functional generative network

Grid

0.25° global

Horizon

15 d, 6-hourly steps

Cadence

00 / 06 / 12 / 18Z

Access

Allow-listed: BigQuery, Earth Engine, Cloud Storage. Mirrored here as a four-field ensemble-mean subset

Licence

Historical data CC-BY-4.0; real-time data under Google DeepMind's experimental data terms

tropics.earth already reads a thousand-member WeatherNext cyclone ensemble as a density rather than as spaghetti, and that page is the argument for this whole site: past a few hundred members, drawing every line is the wrong visualisation.

Live here as a four-field subset of its ensemble mean (2 m temperature, MSLP, precipitation, 500 hPa height), mirrored by Forecast Euro. It is marked experimental everywhere it appears because it is a subset feed, not an operational product. The generation after GenCast, and itself superseded by WeatherNext 3 in August 2026. Its ensemble is cheap enough per member that the useful unit stops being the member and becomes the population — which is a different reading problem, not just a bigger one.

Architecture

Trained on CRPS (formerly AIFS-CRPS): graph-transformer encoder and decoder, sliding-window transformer processor

Grid

0.25° global

Horizon

360 h

Cadence

00 / 06 / 12 / 18Z

Access

Open data from ECMWF, mirrored on AWS, Azure and Google Cloud, same layout as AIFS Single

Licence

CC-BY-4.0

An ensemble a learned model produced itself, rather than by perturbing a physics model. The interesting question is not whether its mean beats the ENS mean — it is whether its spread means the same thing.

Members carried in the same open-data tree as the deterministic run. tropics.earth already decodes its cyclone BUFR through pipelines/ecmwf_bufr.py, so the tracker half of this is proven; the gridded half is not built here yet.

Architecture

Graph neural network on an icosahedral mesh

Grid

0.25° global

Horizon

10 d (240 h in NOAA's feed)

Cadence

00 / 12Z in NOAA's research feed (about 5–9 h behind)

Access

Open weights; also run twice daily by NOAA OAR (noaa-oar-mlwp-data), from both GFS and IFS analyses

Licence

Code Apache-2.0; weights CC-BY-4.0 (commercial use allowed since 6 Aug 2026)

Historical anchor for the verification page. A skill chart that starts at GraphCast shows how fast the field moved.

The ancestor. Not the best model on this page any more, and that is the point of keeping it here — it is the one whose scores everything since has been measured against.

Architecture

3D Swin transformer with a Perceiver encoder, pretrained then fine-tuned

Grid

0.25° global (0.1° HRES fine-tune; 0.4° air quality)

Horizon

240 h in NOAA's feed; no stated maximum rollout

Cadence

00 / 12Z in NOAA's research feed (about 5–9 h behind)

Access

Open weights (Hugging Face); also run twice daily by NOAA OAR from GFS and IFS analyses

Licence

MIT (code and weights)

It is the strongest evidence for the foundation-model claim in this field: that one pretrained atmosphere transfers to problems it was never trained on. /home/maue/ai already has AURO_GFS and AURO_IFS trees, so the initialisation-source question here is ours to answer, not a paper's.

The one that is not only a weather model. Pretrained broadly, then fine-tuned into separate heads — air quality, ocean waves, tropical cyclones — from the same backbone.

Architecture

Spherical Fourier neural operator

Grid

0.25° global

Horizon

10 d

Cadence

00 / 12Z in NOAA's research feed (about 5–9 h behind)

Access

Open weights (NGC, Earth2Studio); also run twice daily by NOAA OAR from GFS and IFS analyses

Licence

Earth2Studio wrapper Apache-2.0; v2-small weights on NGC (licence unconfirmed)

The long-rollout behaviour is the interesting property, not the day-3 score. A model you can spin for a season without it exploding is a different tool from one you can spin for ten days.

The neural-operator branch of the family tree. SFNO learns an operator on the sphere rather than a mapping on a grid, which is why it stays stable when it is rolled out far past its training horizon.

Architecture

3D Earth-specific transformer, hierarchical lead times

Grid

0.25° global

Horizon

Any lead time by chaining 1/3/6/24 h models; 240 h in NOAA's feed

Cadence

00 / 12Z in NOAA's research feed (about 5–9 h behind)

Access

Open weights; also run twice daily by NOAA OAR from GFS and IFS analyses

Licence

Weights CC BY-NC-SA 4.0 — commercial use forbidden

Its hierarchical trick — separate models for separate lead steps, composed greedily — is a reminder that a lot of the early gains came from the training recipe rather than the architecture.

The first learned model to beat the operational IFS on headline deterministic scores, and the one that ended the argument about whether this was going to work.

Architecture

Functional generative network on a mesh transformer

Grid

0.05° at stations · 0.1° surface grid · 0.25° pressure levels

Horizon

15 d at 00/06/12/18Z; 48 h on the hourly interim runs

Cadence

Hourly

Access

Allow-listed: BigQuery, Earth Engine, Cloud Storage (Zarr)

Licence

Google DeepMind experimental data terms for real-time data

The first learned model that starts a run every hour. If that holds up, the question stops being which model is best at day five and becomes how fast a learned forecast can react to what just happened.

Google's current flagship, released August 2026. Forecast Eagle holds one hand-downloaded file of it privately; nothing here is public until the terms allow it.

Architecture

Conditional diffusion over a graph backbone

Grid

0.25° global

Horizon

15 d

Cadence

research / on demand

Access

Open weights

Licence

Code Apache-2.0; weights CC-BY-4.0 (commercial use allowed since 6 Aug 2026)

Diffusion is the reason the spread looks physical instead of blurry. A deterministic learned model optimised for RMSE smooths the field it is least sure about; a diffusion model draws a sharp field and disagrees with itself across members instead.

The model that made the case that a learned ensemble could beat ENS on most headline scores. Its lineage runs straight into WeatherNext.

Architecture

Conditional diffusion, coarse field to kilometre scale

Grid

km-scale regional

Horizon

inherits its driver

Cadence

on demand

Access

Earth-2 NIM

Licence

NVIDIA AI Foundation Models Community License (model); NVIDIA AI Product Agreement (NIM)

The honest framing for every global learned model on this page: they are all coarse. Downscaling is where that gets addressed, and it is generative rather than interpolated — which means the detail is plausible, not measured, and the site has to say so.

Not a forecast model. It takes somebody else's coarse forecast and generates the kilometre-scale detail that a 0.25° grid cannot carry.

Architecture

Diffusion emulator of a km-scale climate model

Grid

km-scale global

Horizon

climate

Cadence

on demand

Access

Open weights on Hugging Face and NGC; code at NVlabs/cBottle. HEALPix, emulating 5 km global output

Licence

Unconfirmed: model card says NVIDIA Software and Model Evaluation License, Hugging Face metadata says Apache-2.0

It breaks the frame. Everything else here answers what happens next; this answers what a Tuesday in this climate looks like — which is a question a forecast cannot be asked.

Climate in a bottle: sample a kilometre-scale atmosphere directly instead of integrating one. Furthest from an operational forecast of anything on this page, and the most interesting thing on it.