There is a great article from the German Meteorological Service on WeatherNext 3 and how AI-based weather models will probably co-exist with physics-based ones.
It's great that they're working on this, but I am puzzled at how absolutely awful the forecasts are in the Google weather app for my area. The forecast will show no rain, the radar view shows nothing, meanwhile it's pouring outside and every other app I check shows it. I know I can't expect it to be perfect, but being terribly wrong even one in ten times is enough to tarnish its reputation permanently.
The Google daily and hourly forecasts aren't great, but the weather map forecast is excellent. Where does that come from? What's generating the predicted movement of the storm clouds?
Weather 2 doesn't seem to have been an ensemble model. Weather 3 is, so theoretically it can get more accurate outcomes by taking the probabilistic analysis of several models concurrently to determine the most likely weather conditions.
I'm building a tool right now that uses an ensemble to determine wind gust likelihood, which is useful for safety critical work on construction sites and the like.
Dark Sky is still one of my all time favorite app. The accuracy of the forecast within the coming hour is to this day unmatched. It constantly reminded me of Back to the Future 2, where Doc says the rain will stop in “in exactly 5 seconds”
The only forecasted rainfall for local areas within a very short timeframe. This made it possible to utilize simple factors to predict the near future.
It is built into Apple Weather after Apple purchased them.
- Units are imperial rather than metric (i.e. not basing it on users locale)
- The concept of init time is not initially clear at all, so it's confusing when you click the calendar icon to see the weather forecast for a future date only to see it isn't an option, you have to use the slider.
- When changing the sidebar to the "detailed" view, most of the hovering element is cut off by the container so you can't actually read it.
And it took longer to write this comment than it did to find these issues. It's fine since it's clearly marked as experimental, but I disagree that it's "really easy to use" :p
Negative. Clicking on any appearance of "Try WeatherNext 3" brings me to this nonsense:
WeatherNext 3 is being integrated into the Google products and tools that billions of people rely on – like Search, Maps, and Gemini. The model is also available for enterprise use, across a range of different applications.
Gain access to high-resolution forecasts, without model setup. Including real-time operational data and historical forecasts.
Try WeatherNext 3 in BigQuery, Earth Engine, Google Maps Platform and Google Cloud Storage.
---
None of those words are links or demos. It's just noise.
I had a problem, and that problem is resolved. Thank you for all of your kindness and assistance on this matter. My gratitude for your effort extends beyond all imaginable boundaries.
No worries, glad to help pitch in for all the times someone has helped me find something obviously staring me in the face :). I swear "I can't find my keys"->"they're in your pocket" type things are a mandatory human experience (if given enough time).
> WeatherNext 3 will power weather features in Google Search, the Gemini app and Google Maps, as well as the Google Maps Weather API and Google Earth Engine.
The online viewer could really, really use Wind Direction as a compass bearing. Its super important considering wildfire/bushfire, air quality, ocean-going conditions, and a myriad of other things. It is produced as a set of vectors during model creation, so would be very useful to see.
In the energy world, this should be such a boon over the classic NWP (Numerical Weather Prediction; complex ML models), but I've not seen it implementated. Anyone with experience of these models over classic NWP?
> Problem is that the WN3 grid is still quite rough (5km) - but a that's a brutal improvement for many places compared to many other global models.
That's a pretty apples-and-oranges comparison. One would almost always use a high-resolution regional model if you needed certain details for different forecasting applications like renewable energy.
It's also worth noting that the 5km outputs are from a model decoder head that was trained against temperature and dewpoint at surface stations. According to the Rasp et al (2026) preprint, this head was designed for continuous sampling; the choice of a 5km grid is arbitrary. What we don't actually know is how well the model handles shocks like a frontal passage or impacts from things like outflow from storms - or even evaporative cooling from precipitation. We are limited to the output that DeepMind publishes; we can't run the model and stress test these things on our own.
That's all a long way to say that the 5km resolution is (a) limited to temperature fields, and (b) we don't know if the "additional" resolution has any impact whatsoever on the phenomena that one would typically use a mesoscale-resolving forecast for.
This would not likely be a great idea since you reduce your ability to understand inputs except for a few parameters. Explainable inputs become very important for many down the line processes used by government and industry alike, because said inputs and their predictive certainty can be quite informative, even critical, for accurate mesoscale prediction.
If model members were available, with all the usual measures, thats a fantastic place to start looking at serious inclusion. It doesnt seem thats the case, however.
TLDR: the input data for the model now also includes real-time observations (satellite and weather stations) on top of the typical (re)analysis data, improving model resolution, run frequency and timestep frequency.
https://www.dwd.de/DE/wetter/thema_des_tages/2026/9/6.html (German only)
I'm building a tool right now that uses an ensemble to determine wind gust likelihood, which is useful for safety critical work on construction sites and the like.
As a former Googler, I wouldn't at all be surprised if this is an integration that is "planned" -- but just not done yet.
And some good handful of people are planning to wring a promo out of work. "Implemented weather UI in Android that is 63% more accurate." ;)
I guesstimate that it has less than 50% accuracy for my area
* https://apnews.com/article/weather-forecasts-worsen-doge-tru...
* https://www.independent.co.uk/news/world/americas/us-politic...
A good book on the history of forecasting, The Weather Machine: A Journey Inside the Forecast:
* https://www.andrewblum.net/the-weather-machine-2
It is built into Apple Weather after Apple purchased them.
- Time is UTC rather than local by default.
- Units are imperial rather than metric (i.e. not basing it on users locale)
- The concept of init time is not initially clear at all, so it's confusing when you click the calendar icon to see the weather forecast for a future date only to see it isn't an option, you have to use the slider.
- When changing the sidebar to the "detailed" view, most of the hovering element is cut off by the container so you can't actually read it.
And it took longer to write this comment than it did to find these issues. It's fine since it's clearly marked as experimental, but I disagree that it's "really easy to use" :p
Link?
edit: The link for the demo is as thus, https://deepmind.google.com/science/weatherlab
Here's a URL for the demo for those who -- you know -- like to click on links and see stuff happen: https://deepmind.google.com/science/weatherlab
404. That’s an error.
The requested URL was not found on this server. That’s all we know.
WeatherNext 3 is being integrated into the Google products and tools that billions of people rely on – like Search, Maps, and Gemini. The model is also available for enterprise use, across a range of different applications.
Gain access to high-resolution forecasts, without model setup. Including real-time operational data and historical forecasts.
Try WeatherNext 3 in BigQuery, Earth Engine, Google Maps Platform and Google Cloud Storage.
---
None of those words are links or demos. It's just noise.
(I'm not trying to be pedantic, but if someone is having trouble finding the button, the exact text is helpful.)
Also, here's where the button takes you: https://deepmind.google.com/science/weatherlab
https://i.imgur.com/IVv4y0n.png
Keep in mind both button are "the demo". One is for trying out the API yourself, the other is for seeing a pre-made dashboard.
> WeatherNext 3 will power weather features in Google Search, the Gemini app and Google Maps, as well as the Google Maps Weather API and Google Earth Engine.
https://dataconomy.com/2026/09/04/weathernext-3-ai-forecasts...
So I assume the main way would be googling "weather Los Angeles" and it will be powered by the WeatherNext 3 models
They also open sourced the last one and are doing B2B/enterprise arrangements so maybe other weather apps are experimenting with it.
Problem is that the WN3 grid is still quite rough (5km) - but a that's a brutal improvement for many places compared to many other global models.
Quite a few country-scale models go down to a 1-2km grid nowadays. This is very helpful in complex geography like mountains and alleys.
That's a pretty apples-and-oranges comparison. One would almost always use a high-resolution regional model if you needed certain details for different forecasting applications like renewable energy.
It's also worth noting that the 5km outputs are from a model decoder head that was trained against temperature and dewpoint at surface stations. According to the Rasp et al (2026) preprint, this head was designed for continuous sampling; the choice of a 5km grid is arbitrary. What we don't actually know is how well the model handles shocks like a frontal passage or impacts from things like outflow from storms - or even evaporative cooling from precipitation. We are limited to the output that DeepMind publishes; we can't run the model and stress test these things on our own.
That's all a long way to say that the 5km resolution is (a) limited to temperature fields, and (b) we don't know if the "additional" resolution has any impact whatsoever on the phenomena that one would typically use a mesoscale-resolving forecast for.
If model members were available, with all the usual measures, thats a fantastic place to start looking at serious inclusion. It doesnt seem thats the case, however.
TLDR: the input data for the model now also includes real-time observations (satellite and weather stations) on top of the typical (re)analysis data, improving model resolution, run frequency and timestep frequency.