Who weathernext is for#
Atmospheric science researchers prototyping AI forecasting methods
WeatherNext 2 gives you a state-of-the-art baseline with full training code, so you can experiment with architecture changes, loss functions, or ensemble methods without reimplementing the entire pipeline. The JAX codebase is readable and modular, with shared utilities for autoregressive rollouts, normalization, and graph building. Use it to test new ideas against a published benchmark.
Skip if:
You need a black-box API for production forecasts without modifying the model. The value here is research flexibility, not plug-and-play deployment.
Developers building weather-dependent applications
If you are building flight routing, agricultural planning, renewable energy forecasting, or disaster response tools, WeatherNext gives you direct access to the same forecasts NOAA used operationally for hurricane tracking. Run backtests over historical data, customize resolution or output variables, and avoid per-API-call costs that scale unpredictably with user growth.
Skip if:
Your application needs real-time forecasts updated every hour with zero infrastructure management. The managed feeds (Google Cloud, OpenMeteo) are better for that use case, though they reintroduce vendor dependency.
Meteorology teams evaluating AI vs physics-based models
WeatherNext 2 is the first AI model with documented operational use in tropical cyclone forecasting (2025 Atlantic hurricane season). If your team is researching whether AI models are ready for operational deployment, you can run side-by-side comparisons against your existing HRES or GFS pipelines using the same evaluation metrics from the Nature paper.
Skip if:
You need ensemble probabilistic forecasts for uncertainty quantification. WeatherNext 2 is deterministic (single forecast per run). The legacy WeatherNext Gen (GenCast) model in this repository does ensembles, but it is an older architecture.
Infrastructure teams training large AI models on atmospheric data
The repository includes the full training loop, gradient computation, and loss functions used in the published papers. If you have access to TPU clusters and want to fine-tune on regional datasets (higher resolution over a specific continent, or extended lead times), the code is production-ready and the WeatherBench2 data pipeline is documented.
Skip if:
You do not have multi-TPU or multi-GPU training infrastructure. Training WeatherNext 2 from scratch or fine-tuning on large datasets requires the same scale of compute as the original DeepMind training runs.
The problem it solves#
Traditional weather forecasting depends on physics-based numerical models that simulate atmospheric equations on supercomputers. The European Centre for Medium-Range Weather Forecasts (ECMWF) operates HRES, the global standard for medium-range predictions, but each forecast run takes hours of computation on massive clusters. Only national meteorological agencies can afford this infrastructure. Researchers, startups, and development teams building weather-dependent applications must rely on commercial APIs (Tomorrow.io, Weatherstack, AccuWeather) or wait for public data feeds that may not match their resolution, cadence, or coverage needs. API costs scale steeply for high-frequency or high-resolution access. For reproducible forecasts, custom model tuning, or offline analysis, neither commercial APIs nor waiting for public feeds is a viable path. You either pay for access you do not control, or you do not get the forecast flexibility you need.
How it solves it#
0.25° resolution global forecasts
Generates medium-range (up to 10 days) atmospheric predictions at roughly 30km resolution, covering temperature, wind speed (including 100m wind for renewable energy applications), geopotential height, and precipitation. The model outputs match the spatial detail of ECMWF's operational HRES forecasts, which is the benchmark for global weather services.
Tropical cyclone track prediction
Includes a direct tracker that extracts cyclone positions, intensities, and tracks from model outputs. The WeatherNext Cyclones variant was used operationally during the 2025 Atlantic hurricane season as FNV3, with NOAA's National Hurricane Center postprocessing it as GDMI. The paper documents breakthrough accuracy in cyclone forecasting compared to traditional models.
Pre-trained operational weights
Ships with multiple model checkpoints trained through 2024 on ECMWF HRES data, including the WeatherNext2_<2025 variant fine-tuned for operational use. Weights are hosted on Google Cloud Storage and downloaded automatically by the inference code, so you do not need to train from scratch.
JAX-based autoregressive rollout
Inference is implemented in JAX and optimized for TPU, with support for GPU via attention implementation switching. The autoregressive rollout generates forecasts step by step, with each 6-hour prediction feeding into the next. The demo Colab notebook shows how to initialize state from HRES initial conditions and run a full 10-day forecast.
Reproducible research code
The repository includes the full training loop, loss computation, and gradient step code used in the published papers (Nature 2026 for cyclones, arXiv:2506.10772 for WN2). You can compute training loss on your own data, fine-tune on regional datasets, or experiment with architecture changes. ERA5 and HRES training data are available via WeatherBench2.
Lightweight mini models for prototyping
WeatherNext Cyclones Mini runs at 1° resolution (roughly 110km) and fits on a free Colab TPU v5e-1 or a single P100 GPU, making it accessible for local testing without H100-class hardware. The mini model forecasts the same variables as the full version, including cyclones, though with reduced accuracy documented in the paper appendix.
Strengths and trade-offs#
Strengths
- Operational-grade accuracy with research transparencyWeatherNext 2 is not a demo. The WeatherNext Cyclones <2025 checkpoint ran live during the 2025 Atlantic hurricane season and was postprocessed by NOAA's National Hurricane Center. The Nature paper documents its performance against ECMWF and other operational models. Unlike commercial APIs that hide their methods, you get the full training code, architecture details, and evaluation benchmarks.
- Apache 2.0 license with no usage restrictionsThe code and notebooks are Apache 2.0 licensed, and other materials (documentation, examples) are CC BY 4.0. You can run unlimited forecasts on your infrastructure, modify the model, and deploy it commercially without licensing fees. Compared to commercial weather APIs that charge per call or tier, the cost model is purely your compute, which scales predictably.
- TPU-optimized inference in minutes, not hoursA 10-day forecast rollout completes in minutes on TPU v5p or H100, compared to hours for ECMWF HRES on supercomputers. For research workflows that need hundreds of forecasts (ensemble generation, backtesting, sensitivity analysis), the speed difference is the difference between viable and impractical.
- Active development by Google DeepMind and Google ResearchThe repository is actively maintained, with the last push on 2026-08-11 (one day before this analysis). It includes legacy models (GraphCast, GenCast published as WeatherNext 1 Graph and Gen) alongside the latest WN2, showing a sustained research program. The team publishes in Nature and maintains public data feeds via Google Cloud and OpenMeteo.
Trade-offs
- -Requires TPU or H100 GPU for full operational modelsThe full WeatherNext2_<2025 and WeatherNext Cyclones models need H100 GPUs or TPU v5p for sufficient VRAM. The mini models run on P100 or free Colab TPUs, but with reduced accuracy. If you do not have access to these accelerators, you are limited to the managed data feeds (Google Cloud, WeatherLab, OpenMeteo) rather than self-hosted inference, which defeats the vendor-independence advantage.
- -Research code with no API stability guaranteesThe README explicitly states this is research code provided as-is, with no guarantees of API stability and breaking changes possible without notice. If you build production workflows on top of WeatherNext, you own the maintenance burden when the repository updates. The recommendation is to pin to a specific release, which means you miss bug fixes and model improvements unless you actively track and merge updates.
- -Requires atmospheric data expertise to validate outputsWeatherNext outputs raw atmospheric variables (geopotential height, wind components, temperature, humidity). Interpreting these for real-world decisions (should this flight route avoid a storm, is this harvest window safe) requires domain knowledge or postprocessing pipelines. The cyclone tracker helps for tropical systems, but for general use cases you need to build your own validation and alerting on top of the model outputs.
- -Training data governed by separate ECMWF termsFine-tuning or retraining requires ERA5 and HRES data from ECMWF, accessed via WeatherBench2. These datasets have their own terms and conditions (ECMWF open data is CC BY 4.0, but check current restrictions). If you cannot comply with ECMWF's data policies, you cannot retrain the model on the same distribution it was validated against.
weathernext vs alternatives#
WeatherNext 2 vs Commercial Weather APIs
Commercial APIs (Tomorrow.io, Weatherstack, AccuWeather) provide forecast data via REST endpoints with per-call or subscription pricing. WeatherNext 2 is an open source model you run on your infrastructure with one-time compute costs.
| Feature | WeatherNext 2 | Commercial APIs |
|---|---|---|
| License | Apache 2.0 | Proprietary |
| Hosting | Self-hosted TPU/GPU | Managed cloud |
| Resolution | 0.25° (30km) | Varies |
| Cost | Compute per run | Per-call fees |
| Cyclone tracking | Included | Premium tier |
| Customization | Full model access | API params only |
WeatherNext 2 wins for reproducible forecasts, high-frequency access, and avoiding vendor lock-in. A startup running 10,000 queries daily can pay thousands monthly for APIs; self-hosted WeatherNext costs a fixed hourly rate. The Apache 2.0 license allows modification, fine-tuning, and commercial integration without restrictions.
Commercial APIs are better when you need zero infrastructure management, real-time updates, or hyperlocal nowcasting (WeatherNext focuses on 1-10 day forecasts). For low-volume widgets, managed simplicity outweighs cost. WeatherNext's value is for teams with ML infrastructure who need forecast pipeline control.
WeatherNext 2 vs ECMWF HRES
ECMWF HRES is the global standard for medium-range forecasting, a physics-based numerical model on supercomputers. WeatherNext 2 is an AI model trained to match HRES accuracy but run faster.
| Feature | WeatherNext 2 | ECMWF HRES |
|---|---|---|
| Model type | AI neural network | Physics-based |
| Resolution | 0.25° (30km) | 0.1° (9km) |
| Forecast time | Minutes (TPU/GPU) | Hours (supercomputer) |
| Accessibility | Open source | Public feeds only |
| Operational use | NOAA 2025 cyclones | Global standard |
| Transparency | Full code/weights | Closed |
WeatherNext 2 replicates HRES outputs via fine-tuning on HRES data. The Nature 2026 paper shows comparable cyclone accuracy, which led to NOAA adoption in 2025. Speed (minutes vs hours) makes ensemble studies practical for teams without supercomputer access.
HRES is better for operational meteorology needing interpretability. Physics-based failures are debuggable; WeatherNext is a black-box neural network. HRES runs higher resolution (9km) with live data assimilation; WeatherNext initializes from HRES outputs. For public weather warnings, HRES remains the standard. For research and custom applications, WeatherNext offers comparable accuracy with lower compute and full transparency.
vs Legacy Models (GraphCast, GenCast)
GraphCast (Science 2023) is deterministic forecasting via graph neural networks. GenCast (Nature 2024) is diffusion-based ensemble forecasting. WeatherNext 2 (2024 training) is optimized for operational deployment, using HRES operational data instead of only ERA5 reanalysis. It adds 100m wind prediction and a direct cyclone tracker. The operational checkpoints were used by NOAA in 2025.
Use GraphCast for lightweight deterministic baselines. Use GenCast for ensemble probabilistic forecasts with uncertainty. Use WeatherNext 2 for operational-grade forecasts, cyclone tracking, or renewable energy wind analysis.
Install and self-host#
Installation requires Python and pip. The package is not on PyPI, so install directly from the GitHub repository. Pin to a specific release tag (e.g., v0.3.0) to avoid breaking changes, as the README notes this is research code with no API stability guarantees.
```bash
pip install git+https://github.com/google-deepmind/[email protected]
```
For interactive experimentation, the Colab notebook is the fastest path. Open it from Colaboratory at the link below, select a TPU v5e-1 runtime (free tier), and run all cells. The notebook downloads pre-trained weights automatically and includes sample HRES initial conditions.
Colab link: https://colab.research.google.com/github/google-deepmind/weathernext/blob/master/docs/weathernext2/wn2_demo.ipynb
For local inference on GPU instead of TPU, the notebook shows how to switch the attention implementation. Full operational models need H100 GPUs; mini models work on P100. Model weights are fetched from Google Cloud Storage (https://console.cloud.google.com/storage/browser/dm_graphcast) during the first run.What it's built on#
- Languages
- Python
FAQ#
Is WeatherNext 2 free to use for commercial applications?
Yes. The code and notebooks are Apache 2.0 licensed, which allows commercial use, modification, and distribution with no restrictions. You can run WeatherNext 2 on your infrastructure and use the forecasts in paid products. The only constraint is that training data (ERA5, HRES) from ECMWF has separate terms (typically CC BY 4.0 for open data), so check those if you are retraining or fine-tuning. Pre-trained model weights are provided and covered by the Apache 2.0 license.
Can I run WeatherNext 2 on a laptop or do I need cloud GPUs?
Oops! Something went wrong
[next-mdx-remote-client] error compiling MDX: Unexpected character `2` (U+0032) before name, expected a character that can start a name, such as a letter, `$`, or `_` More information: https://mdxjs.com/docs/troubleshooting-mdx
How does WeatherNext 2 compare to ECMWF HRES for forecast accuracy?
Oops! Something went wrong
[next-mdx-remote-client] error compiling MDX: Unexpected character `2` (U+0032) before name, expected a character that can start a name, such as a letter, `$`, or `_` More information: https://mdxjs.com/docs/troubleshooting-mdx
Do I need to download weather data myself or are initial conditions included?
The repository includes sample HRES initial conditions in the Colab notebooks to get started immediately. For operational use, you need to download HRES initial conditions (for the operational models) or ERA5 data (for the research checkpoints) from ECMWF via WeatherBench2. The demo notebook shows how to load these formats. Pre-trained weights are automatically downloaded from Google Cloud Storage when you run the inference code, so you do not need to fetch those separately.
What is the difference between WeatherNext 2 and the older GraphCast/GenCast models?
GraphCast (now called WeatherNext 1 Graph) is a deterministic forecasting model using graph neural networks, published in Science 2023. GenCast (WeatherNext 1 Gen) is a diffusion-based ensemble model for probabilistic forecasts, published in Nature 2024. WeatherNext 2 is the latest architecture, trained through 2024 and fine-tuned on operational HRES data for direct initialization from live forecasts rather than reanalysis. All three models are in this repository, but WeatherNext 2 is the current operational version and the one used by NOAA in 2025.
Similar open-source tools#
PyScrappy
AI-native web scraping toolkit + MCP server
agent-toolkit-for-aws
Empower AI agents to build and manage AWS applications
deer-flow
Build super agents with DeerFlow's powerful framework
Anthropic-Cybersecurity-Skills
Empower AI agents with structured cybersecurity skills
daily_stock_analysis
AI-driven stock analysis for A, HK, and US markets
timesfm
Foundation model for time-series forecasting