Your weather app is about to get a lot smarter, whether Google makes it or not. The company’s new model, WeatherNext 3, is the most accurate AI forecaster ever tested. Here’s how it works, minus the jargon.
What is WeatherNext 3?
WeatherNext 3 is an AI model built by Google DeepMind and Google Research that predicts temperature, rain, wind, and humidity. Instead of solving physics equations the traditional way, it learned weather patterns from more than fifty years of historical data released by the European Center for Medium-Range Weather Forecasting, known as ECMWF.
Here’s the quick profile:
- Who built it: Google DeepMind and Google Research
- What it predicts: temperature, rain, wind, and humidity
- How it learned: more than 50 years of ECMWF weather data
- Where it ranks: first among AI forecasters on Operational WeatherBench
The results are hard to argue with. Tested on Operational WeatherBench, a standard comparison tool built by the startup Brightband, WeatherNext 3 beat every other AI model in the running, including ones from Microsoft, Nvidia, and the ECMWF itself. More importantly, it also beat traditional forecasts from the US National Weather Service and the ECMWF, the systems we’ve trusted for decades. TechCrunch has the full breakdown, including the quotes from the Google engineers who built it.
For a beginner, the short version is: the free weather data on your phone is about to get more accurate than the government supercomputers that have been setting the standard for years.
Why AI weather models are beating the old ones
Traditional forecasting is a miracle of engineering, and also kind of a dinosaur. It works like this: government agencies run giant supercomputers that crunch through mathematical equations describing the physics of the atmosphere. Those systems are remarkably accurate, but they’re expensive to run and slow to update, which is why most forecasts refresh every few hours.
AI models flip the approach. Instead of solving equations, they study decades of past weather data and learn the patterns: how pressure systems move, how rain forms, how wind follows temperature. Once trained, a model can produce a forecast in seconds, at a fraction of the cost. It’s the same pattern we’re seeing across AI right now, and it’s why news like the Nvidia and Hugging Face deal matters: the weight of these models is shifting to whoever can train them on real data.
The catch is that earlier AI models had real weaknesses. They forecast over wide areas, 15 to 25 square kilometers, which is too broad to tell you whether your street gets rain. They weren’t great with precipitation. And they still depended on data formatted by government agencies. WeatherNext 3 attacks all three problems. Google’s previous model, WeatherNext 2, was already competitive with the old systems; the new version closes the gaps that kept AI forecasts second-guessed by meteorologists.
What’s actually better this time
Google says the new model makes three concrete jumps, and they’re worth breaking down side by side:
| Problem with earlier AI forecasts | How WeatherNext 3 fixes it |
|---|---|
| Forecasts covered 15-25 sq km, too broad to be useful for daily plans | Predicts down to a 5 km resolution |
| Weak at predicting rain, the thing people check most | Rain evaluations improved 60% over WeatherNext 2 |
| Updates every 6 hours, easy to be caught off guard | Produces hourly forecasts |
The rain improvement matters more than it sounds. Most people glance at a forecast for one question: will it rain where I am? Earlier AI models were genuinely shaky there, which is why meteorologists stayed skeptical. A 60% improvement on rain accuracy is the kind of number that changes minds.
The resolution jump is the other big one. A 5 km forecast is roughly neighborhood-level, not city-level. It’s the difference between “chance of showers” and “bring an umbrella around 4 p.m., the block near the park gets hit.” For planning a run, a commute, or an outdoor event, that’s a real upgrade.
Where you’ll see it
Google says WeatherNext 3 will feed into the weather info shown in Google Search, Google Maps, and Gemini.
In Google products
A senior staff engineer at Google told TechCrunch that this is the first time some of these core variables will power a lot of Google products at once. For users, that means weather answers in Search get sharper, and asking Gemini about weekend plans will pull from a better forecast.
For researchers and businesses
Researchers and businesses get access too, through Google Cloud. That part matters even if you never touch it: companies use weather models for everything from delivery routes to energy grids, and a better model upstream means better decisions downstream.
And if you don’t use Google at all? You still win. Weather forecasters everywhere benchmark against the best available model. When one lab pushes accuracy up, the whole field pushes to catch up. The rain forecast in your favorite weather app has a good chance of improving over the next year, regardless of who makes it. It’s the same dynamic you see with free open-weight models like GLM 5.3: one strong release raises the bar for everyone, including the tools you already use.
What it still can’t do
For all the hype, AI weather forecasting has hard limits, and it’s worth knowing them before you trust it with your weekend.
First, the models still depend on government data. WeatherNext 3 doesn’t observe the sky itself. It learned from the decades of records that agencies like ECMWF and the National Weather Service produced, and it keeps needing fresh input from those systems to make predictions. If a government agency stopped collecting weather data, every AI forecaster would go blind. The AI revolution in weather is an upgrade, not a replacement.
Second, chaos sets a ceiling. Weather is famously chaotic: tiny differences in the starting conditions grow into big differences later on. That’s why forecasts get fuzzy beyond a week or two, no matter how good the model is. AI has made the short and medium range, the next few days, much sharper. It hasn’t cracked the two-week wall, and the physics of the atmosphere says it probably never will.
Third, benchmarks test everyday weather, not every extreme. Operational WeatherBench measures things like temperature, wind speed, and humidity, the variables you check before leaving the house. Sudden severe weather, the kind that turns into a warning siren, still leans on human forecasters and official alert systems. The stakes there are too high to trust a probability alone.
None of that is a knock on WeatherNext 3. It’s the honest way to read the numbers: this is the best AI weather model ever tested, and even the best model is one part of a bigger forecasting system.
The takeaway
WeatherNext 3 is the best-tested AI weather model to date: sharper resolution, dramatically better rain predictions, and hourly updates instead of six-hour gaps. It’s already flowing into Google products and will nudge the entire forecasting industry forward. The only thing it can’t do is make you remember your umbrella.
That part is still on you.
