Running a weather model and reproducing a published cyclone evaluation are different tasks. The checkpoint, input data and evaluation period must match the question.
A stronger forecast is still a forecast
DeepMind’s August 6 announcement describes retrospective 2023-2024 evaluation and use by the US National Hurricane Center in 2025. Its average lead-time improvement does not mean every cyclone gives every community an extra day to evacuate. The announced 2026 ensemble grows to 1,000 members; that count is not a guarantee that every possible outcome is represented.
For public decisions, follow warnings from the responsible meteorological agency. A locally generated trajectory is research output, not a replacement warning service.
Match training history to the test year
The official repository documents separate Cyclones checkpoints: the one trained through 2022 supports evaluation on 2023; through 2023, evaluation on 2024; through 2024, operational use in 2025. Using later training data to claim a clean historical forecast would answer a different question.
Mini uses a 1-degree grid, versus 0.25 degrees for the full model, and does not reproduce the full model’s performance. WeatherNext 2 and Cyclones share the algorithm, but their weights are independently trained. Code and notebooks use Apache 2.0; other repository materials use CC BY 4.0. Input datasets still require their own terms to be checked.
What the grid illustration explains
The cover divides an angular 1-degree square into sixteen 0.25-degree cells. Four subdivisions on each axis give 4 × 4 = 16. This is a grid illustration, not a claim that the finer model is sixteen times more accurate. Physical cell width also varies with latitude, and model skill depends on more than spacing.
A smaller model can be useful for learning the pipeline and checking that inputs, coordinates and output files are handled correctly. Treat a successful demonstration as that result, rather than as a reproduction of every number in a paper.
Save enough information to reproduce the run
Record checkpoint name, training cutoff, repository revision, initialization time, input dataset version and ensemble settings. Keep track extraction and verification code alongside the output. Changing post-processing can change a track score even when the underlying forecast fields are identical. That record makes a later comparison interpretable; a screenshot of an impressive storm path does not.
September 8: distinguish cyclone checkpoints and training cutoffs, full versus Mini resolution, research evaluation and official warnings; remove unsupported universal lead-time claims.