FGN

Closed weights Google DeepMind 720M parameters June 2025

No estimate

No hardware requirements for this model

The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation
Google DeepMind
Organisation type
Industry
Country
United States of America
Published
12 June 2025
Authors
Ferran Alet, Ilan Price, Andrew El-Kadi, Dominic Masters, Stratis Markou, Tom R. Andersson, Jacklynn Stott, Remi Lam, Matthew Willson, Alvaro Sanchez-Gonzalez, Peter Battaglia

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
Earth science
Task
Weather forecasting

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters
720M

"In this work we train 𝐽 = 4 models. To ensemble their predictions, we generate an equal number of ensemble member trajectories from each model" "FGN is a larger model, with ∼180m parameters per model seed" They train a 4 model ensemble with 180M parameters per model. Total parameters: 4*180=720M

Training data
tokens

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute
9.6 × 10²¹ FLOP

TPU v5 FLOPs: 459000000000000 TPU v6 FLOPs: 918000000000000 They train for 490 TPU days, with an unspecified mix of v5 and v6. Assuming same proportion of both TPU and a utilization of 0.33. Mean TPU FLOPs: 688500000000000 Compute: 490*24*60*60*688500000000000*0.33=9618950880000001000000

How it was established
Hardware

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware
Google TPU v5p,Google TPU v6e Trillium
Wall-clock time
72 hours

"Training FGN takes approximately 3 wall clock days, using a combined total of 490 TPUv5p and TPUv6e days of compute. " It's unclear if training a single model or the whole ensemble takes 3 days.

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights
Closed — provider access only
Model access
Unreleased
Training code
Unreleased

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Why it is tracked
SOTA improvement

"This paper presents FGN, a simple, scalable and flexible modeling approach which significantly outperforms the current state-of-the-art models" "FGN achieves state-of-the-art cyclone track prediction" "Our results show FGN offers substantial improvements over previous ML-based probabilistic weather models, and sets a new state-of-the-art in ensemble forecasting. It outperforms GenCast with both the skill of its marginal forecasts, including on extreme weather, as well as its skill in forecast…

Record confidence
Confident

Sources

Where this record came from and when it was last checked.

Reference
Skillful joint probabilistic weather forecasting from marginals
Last updated
28 November 2025

What the numbers mean

Background

FGN was published by Google DeepMind, in United States of America, in June 2025. The organisation is categorised as industry.

It works in Earth science, and is recorded as doing weather forecasting.

This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.

What went into building it

Producing it required around 9.6 × 10²¹ FLOP of arithmetic, on Google TPU v5p,Google TPU v6e Trillium, which is a statement about the training budget rather than about inference.

Its inclusion criterion is sOTA improvement.

Answers

FGN — common questions

01

Is FGN open source?

No. FGN has not had its weights published, so it exists only as a service controlled by its owner.

02

How many parameters does FGN have?

FGN has 720M parameters. "In this work we train 𝐽 = 4 models. To ensemble their predictions, we generate an equal number of ensemble member trajectories from each model" "FGN is a larger model, with ∼180m parameters per model seed" They train a 4 model ensemble with 180M parameters per model. Total parameters: 4*180=720M. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

03

Who created FGN?

FGN was published by Google DeepMind, based in United States of America, categorised as industry.

04

When was FGN released?

FGN was published in June 2025.

05

What is FGN used for?

FGN works in Earth science, and is recorded as handling weather forecasting. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

06

How much compute was used to train FGN?

Around 9.6 × 10²¹ FLOP, on Google TPU v5p,Google TPU v6e Trillium. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

07

What GPU do I need to run FGN?

None. FGN is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.

Source

Original publication

Record last updated 28 November 2025

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.