Google's WeatherNext 3 Tops Forecasting Benchmarks, Will Power Search and Maps

Google DeepMind and Google Research just shipped WeatherNext 3, and the benchmark numbers are hard to argue with. On Operational WeatherBench, an independent testing platform built by startup Brightband, the model beats every other AI weather system on the leaderboard, including entries from Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting (ECMWF). It also outperforms the traditional physics-based forecasts from the U.S. National Weather Service and ECMWF itself.

That last part matters more than it might seem. Traditional forecasting runs on government supercomputers grinding through differential equations that model atmospheric physics. It works, and it's gotten remarkably good over decades of refinement. But it's slow and expensive to run at scale. Deep learning models trained on historical weather data can approximate that same physics far more cheaply, which is exactly why this space has exploded since ECMWF released 50-plus years of reanalysis data back in 2018.

Google says WeatherNext 3 will start feeding directly into products people already use: search, Google Maps, and Gemini, plus availability for researchers on Google Cloud. "This is going to be the first time that some of the core variables feed and power a lot of the Google products," Samier Merchant, a senior staff engineer at Google, told TechCrunch. In other words, the next time you check the weather in a Google product, there's a decent chance this model is doing the work behind the scenes.

What Actually Changed Under the Hood

Three known weaknesses have dogged AI weather models since the field took off: they tend to forecast at coarser resolution than is genuinely useful (typically 15 to 25 square km grids), they struggle with precipitation specifically, and they lean heavily on pre-processed government datasets rather than raw sensor data. WeatherNext 3 goes after all three.

• Resolution: down to 5 km on key variables, a meaningful jump from the 15-25 km range that's standard for the category.
• Precipitation accuracy: a 60% improvement over WeatherNext 2 on rain evaluations, which is the metric AI models have historically been worst at.
• Update frequency: hourly forecasts instead of the standard six-hour cadence most models, AI or physics-based, still use.

None of that is free. WeatherNext 3 is 2.4 times larger than its predecessor in parameter count, and the team retuned the targets for its decoder heads (the output layers that translate internal representations into specific predictions) to produce more actionable results. DeepMind has also leaned into task-specific outputs before, winning praise for visualizing cyclone paths directly instead of just spitting out grid-averaged metrics.

The bigger architectural shift is training the model to forecast conditions at specific weather stations, not just across a 3D grid. "Adding a capability where this model is now also predicting, say, what Denver's airport's weather station is going to measure on an hourly basis, just connects that forecasting task closer to the core," said Daniel Rothenberg, an atmospheric scientist at Brightband. It's a practical move too: station-level predictions are easier to validate against ground truth, since you can directly compare the model's output to what a sensor actually recorded.

The hourly update capability comes from a specific data choice: WeatherNext 3 ingests real-time satellite data rather than relying solely on the batch-processed analysis products that government weather centers produce on longer cycles. Google is calling this the first AI model to directly incorporate raw observations for a high-resolution global forecast. That claim didn't go unchallenged. WindBorne, an AI weather startup running its own weather balloon fleet, says its WeatherMesh 6 model has been ingesting raw observations since late 2025. Google's response was that its forecasts run at higher resolution globally. Worth noting: both models still lean on national weather datasets for parts of their pipeline, so full end-to-end data assimilation, skipping government-processed data entirely, isn't quite here yet.

Ferran Alet, a staff research scientist manager at DeepMind, framed the underlying appeal of the approach simply. "Weather is chaotic, and so small differences really start to perturb massively," he said. "Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data." That's a decent one-line summary of why transformer-based models have taken over so much of this field so fast: you're not solving the physics exactly, you're learning the statistical patterns that make the physics tractable at scale.

Key Takeaways

The transformer revolution in meteorology doesn't get nearly the attention that LLMs do, but it's arguably just as consequential for real-world impact. European and U.S. weather agencies are already deploying AI models operationally, not experimentally. The economics are the real story here: AI forecasting is cheap enough to bring high-quality predictions to regions that could never afford the sensor networks and supercomputing infrastructure that developed nations take for granted.

Bill Gates recently pointed to AI weather forecasting as one of the technology's clearest near-term benefits, citing improved crop yields in developing countries as a direct downstream effect of better predictions. Alet added that higher-resolution wind, rain, and cloud cover forecasts have a parallel use case in renewable energy, where dependable output forecasting is a real bottleneck for grid operators trying to integrate solar and wind at scale.

For Google, there's an obvious product angle too. "At the end of the day, I think Google is about providing useful information to the user, and a lot of what users are looking for has to do with the weather in some way or another," Alet said. It's a low-key statement for what amounts to a fairly significant infrastructure shift: the weather data behind some of the most-used products on the internet is quietly getting rebuilt on a neural network.