📰 Key Takeaways

This is Shimizuwan’s daily translation task, outputting a Traditional Chinese summary directly.

Google DeepMind and Google Research today unveiled WeatherNext 3, billed as the most advanced and accurate global weather AI model around, ranked best in the industry based on real-time evaluation results from independent group Brightband. The model’s biggest breakthrough is that it learns directly from real-time observation data instead of relying solely on the historical records used by traditional numerical weather prediction (NWP) models — it generates high-resolution forecasts hourly from raw satellite data, making predictions more immediate and more attuned to local weather shifts. On resolution, WeatherNext 3 delivers 5km resolution for key surface variables like temperature and humidity, 10km for other surface variables, and 25km for atmospheric variables like wind speed — roughly five times finer overall than its predecessor WeatherNext 2 (25km grid, updated every 6 hours). Architecturally, the model ingests hourly-updated geostationary satellite mosaic imagery alongside traditional historical analysis data, feeding it into a single, flexible Functional Generative Network (FGN) mesh transformer. Its output includes dense gridded data, discrete hurricane tracks, and native forecasts at sparse weather-station coordinate points. Take the UK’s 2-meter temperature forecast as an example: at 25km (0.25-degree) resolution, WeatherNext 2’s output looks pixelated and over-smoothed, while WeatherNext 3 at native 5km (0.05-degree) resolution clearly captures complex local terrain detail. The model will be integrated into Google’s global product line going forward, delivering more reliable weather forecasting services. See the original article for the full technical details.

💬 JudyAI Lab’s Take

WeatherNext 3 isn’t just another weather model update — the key is that it ditches traditional NWP’s reliance on historical data and instead ingests real-time satellite observations directly, generating forecasts hourly. That signals weather prediction shifting from “retrospective modeling” to “real-time perceptual modeling.”

For AI builders, this is a mental model worth keeping in your toolkit: when your data source shifts from “periodic batch snapshots” to “continuous stream,” your architecture should evolve with it. WeatherNext 3 feeds geostationary satellite imagery and historical analysis data together into a single FGN mesh transformer, rather than training each separately and stitching them together afterward. This “unified input, flexible output” design is what let resolution jump from 25km straight to 5km — 5x more detail. The output also supports three formats at once — dense grids, discrete tracks, and sparse coordinate points — which shows that good architecture should satisfy multiple downstream needs in one shot, rather than reinventing the wheel for each output type.

Next time you’re designing a predictive system, ask yourself: is your model still stuck only ingesting historical snapshots, ignoring the real-time streaming data you might already have on hand?

📅 Source Info

🔗 Further Reading