Google DeepMind put out a blog post this morning about WeatherNext 2. They call it their most advanced weather model yet. I call it another Monday where Google reminds everyone they have a ton of TPUs sitting around doing weather now.
The big change this time is speed and the ability to crank out hundreds of different possible forecasts from one starting point. Old-school physics models on supercomputers take hours to do that. WeatherNext 2 does it in under a minute on a single TPU. Eight times faster overall than whatever they were running before.
Resolution got a bump too. They say it can go down to hourly now instead of the six-hour chunks most global models use. That actually matters if you’re trying to figure out whether the rain hits at 2 p.m. or 5 p.m. and ruins your barbecue plans.
The trick they pulled is something called a Functional Generative Network. Basically they inject noise straight into the math part of the model instead of the usual pixel-style noise you see in image generators. Keeps everything physically plausible while still giving you a bunch of different outcomes. Meteorologists care about that because rare nasty events live in the tails of the distribution.
They tested it against the old WeatherNext and say it beats the previous model on 99.9 percent of variables across all lead times out to fifteen days. Temperature, wind, humidity, pressure levels, the works. I’ll take their word for it until someone else runs the numbers.
Real-world side of this is they already shoved the new model into a bunch of Google products. Search, Gemini, the Pixel Weather app, Maps Platform’s weather layer. In the next few weeks regular Google Maps gets the upgrade too. If you’ve noticed the forecast looking a little different today, that’s probably why.
They also opened it up for the data nerds. WeatherNext 2 output is in Earth Engine and BigQuery right now. If you’re on Vertex AI there’s an early access program where you can run your own inferences. Probably costs a fortune per token like everything else on there, but at least it exists.
One thing I always wonder with these AI weather models is how much they actually help the pros. National weather services still run their own physics ensembles because lives depend on it. Google keeps saying they want to work with agencies, and they’ve done some cyclone stuff experimentally, but I haven’t seen anyone ditch their supercomputers for DeepMind yet.
Still, hourly global coverage that runs stupid fast is legitimately useful for energy trading, logistics, insurance pricing, all that fun stuff. And for the average person it means the ten-day in Google Search might stop being complete fiction past day five.
