The repository is for experiments, while the feeds are for applications
WeatherNext packages three generations of Google DeepMind weather work under one roof: WeatherNext 2, the cyclone variants, GenCast, and GraphCast. The code loads model weights, prepares atmospheric inputs, runs autoregressive forecasts, tracks cyclones, and exposes training pieces. If your product only needs forecast values, the README points elsewhere. Daily outputs are already available through Google Cloud, WeatherLab, and OpenMeteo.
That split should decide whether you clone the 97-file repository. Researchers may need custom initial conditions, gradients, architecture changes, or direct access to the 0.25-degree model. An application showing tomorrow's forecast probably does not. Using a feed removes weight management, accelerator provisioning, and much of the data-shaping work, while the repository gives you control over the model itself.
WeatherNext 2 favors TPUs and reserves the full model for large accelerators
The default notebook uses WeatherNext Cyclones Mini at 1-degree resolution. Its checkpoint is 216 MiB, and the README recommends a free Colab v5e-1 runtime or a P100-class GPU. The full WeatherNext 2 and cyclone models run at 0.25-degree resolution with four 701 MiB seed files. Google recommends v5p TPUs; GPU users need to change the attention implementation and provide an H100 with enough memory.
These requirements make no speed claim. We did not run inference or compare forecast skill in our 3-CPU sandbox. The full Cloud TPU guide starts with a billed Google Cloud project, warns that the $300 trial does not cover TPUs, and tells users to check hourly pricing. It also explains VM deletion because the instance continues billing until removed. A successful Python install does not answer the cost question.
What happened when we ran it
Our sandbox cloned commit f2f2c51 and installed 94 packages in 62 seconds under Python 3.12. The environment consumed 863 MB before any model weights or forecast inputs were downloaded. The package build completed in 9 seconds, and pip-audit found 0 known vulnerabilities in the installed dependencies.
There was no test script or target, so we skipped tests rather than inventing a substitute. Our scan also found 0 CI workflow files, no Dockerfile, and no tests directory. The 10 MB checkout contained about 22,432 lines of source. This proves that packaging completed in a clean Debian container. It does not prove that a checkpoint loads, a TPU rollout completes, or the generated weather fields are correct.
The 863 MB environment is only the entry fee
Pretrained weights live in a public Google Cloud bucket. Inference also needs initial weather state with the variables, levels, timestamps, and resolution expected by the selected checkpoint. The notebook demonstrates that path with HRES-style inputs. Training is a larger undertaking: the README directs users to ERA5 through WeatherBench2 and names HRES data for operational fine-tuning, with separate terms that each user must check.
The package metadata pins xarray at 2026.2.0 or earlier and lists Linux with Python 3.10 and 3.11. Our Python 3.12 lab image installed successfully, but that is a measurement of commit f2f2c51 rather than a documented support promise. The TPU VM guide itself creates a Python 3.12 environment and separately pins pandas below version 3, which shows how closely the notebook depends on its scientific Python environment.
Version 0.3.0 is active research code without a test target
Release v0.3.0 arrived on August 6, 2026 with WeatherNext 2 support. The last push on September 4 added the Cloud TPU VM guide, so the later code activity makes an old-tag abandonment claim indefensible. GitHub showed 7,691 stars, 79 combined issues and pull requests, and 74 open issues on September 28. Many open threads concern input preparation, dependency combinations, hardware setup, or interpreting model outputs.
The README gives the more useful maturity label: research code supplied as-is, with no API-stability guarantee and possible breaking changes without notice. It also says this is not an officially supported Google product. That language, the v0.3.0 version, and the absent test target all point in the same direction. Pin a commit, save exact model and data identifiers, and validate generated fields against your own scientific acceptance checks.
Official warnings remain the authority
Google says the models were not produced with or endorsed by a government meteorological agency. The outputs do not replace official alerts, warnings, or notices. That restriction belongs in the product design, especially for cyclone tracking. A polished map or an 8-member ensemble does not turn an experimental output into an emergency-warning source.
WeatherNext publishes code, weights, notebooks, model history, and a detailed TPU path. Our 62-second install makes inspection easy, but running the model is the expensive part. Choose the repository for experiments that require control over inference or training. Choose the feeds for applications that need forecasts without becoming an AI weather lab.