Google published a new white paper this week detailing how it is opening up the source code and model weights behind several of its disaster forecasting systems, giving outside developers, researchers and national weather agencies direct access to tools the company already uses in production.
The announcement, posted to the official Google blog, frames the releases around a single goal: “Our north star is simple yet ambitious: No one should be surprised by a natural disaster.”
What Google is releasing
The centerpiece of the disclosure is a recap of code Google has pushed to GitHub over the past several months rather than a single new tool.
The company highlighted several of its Earth AI datasets, models and APIs that have been made available outside the company, including WeatherNext, NeuralGCM, its hydrology modeling framework, and its Open Buildings datasets.
Those releases build on work published earlier in 2026. In August, Google DeepMind and Google Research put the code and weights for their cyclone forecasting system on GitHub.
Google DeepMind open-sourced the code and model weights via GitHub under an Apache 2.0 licence across three variants: WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini (capable of running in a free Google Colab notebook).
The smallest variant was built specifically so developers without access to specialized hardware could still experiment with it.
A flood forecasting framework built on PyTorch
In June, Google separately released the code behind its river flood forecasting system on GitHub. The project, now referred to by Google researchers as OpenHydroNet, is a heavily modified fork of the NeuralHydrology repository.
On June 3, 2026, Google open-sourced the hydrology framework that powers the river forecasting models driving its global Google Flood Hub platform. The hydrology model is provided as a Python package built using the open-source PyTorch machine learning library.
The Google Flood Forecasting team developed OpenHydroNet to support R2O and operational use cases, and OpenHydroNet is a heavily modified fork of the NeuralHydrology Github repository.
Google said the goal was to let local scientists plug in their own data rather than rely solely on Google’s infrastructure.
By open-sourcing the architecture, Google allows local scientists and meteorologists to plug their specific regional datasets directly into the framework, and benchmarking demonstrates it extends the reliable predictive horizon by six days in gauged basins and by one day in ungauged basins compared to the original model.
Google developed this framework internally and tested it with partners like the Czech Hydrometeorological Institute (CHMI).
Why Google is opening the code now
The white paper ties the various releases to a broader argument that AI-based crisis response cannot be built by one company alone.
The scale of what Google already operates with these models is part of that pitch. In 2025 alone, Google helped connect people with crisis information more than 10 million times per day, on average.
The paper also details other elements of Google’s disaster technology stack that remain proprietary for now, including a satellite system for wildfire detection.
Near-real-time fire boundary maps are live in 34 countries, and FireSat, a global satellite constellation, is designed to detect and track fires as small as a classroom within 20 minutes once completed.
By contrast, the forecasting models for cyclones and floods are the pieces Google has chosen to hand directly to outside developers as working code rather than keep behind an API.
What outside developers can do with it
Because the repositories are released under an Apache 2.0 license, a permissive terms structure similar to those used elsewhere in cloud infrastructure, including this cooling design, developers and researchers can modify, redistribute and commercialize the code with minimal restriction.
Google frames this as a deliberate tradeoff. The framework is developed to allow hydrological scientists to build on what Google Research has done by adding and testing new models, data, and approaches.
For national weather services without the budget to build comparable AI infrastructure from scratch, similar to this fiber link, the released code offers a shortcut.
Whether more agencies adopt the models at the scale Google hopes will depend on how well the open source versions perform once they are running on infrastructure Google does not control.
