Three data-driven weather models have become the reference points for AI meteorology: Microsoft's Aurora, Google DeepMind's GraphCast, and Huawei's Pangu-Weather. All three are trained on ERA5, all three forecast the global atmosphere in seconds on a single GPU, and all three are competitive with — and in many cases better than — physics- based operational forecasts at medium range. But they are not identical. This post walks through how they differ and what each one is best at.
The three contenders
Aurora
Microsoft's Aurora is framed explicitly as an atmospheric foundation model. A 3D Swin Transformer backbone is pretrained on ERA5 and other atmospheric data, then fine-tuned into multiple heads: medium-range weather forecasting, global air quality prediction, and tropical cyclone tracking. The foundation-model framing is what sets it apart — one backbone, many applications.
GraphCast
Google DeepMind's GraphCast uses a graph neural network over an icosahedral grid of the globe. Each grid point is a node, and edges connect neighboring points at multiple resolutions. The model iterates the grid forward in time and predicts next-state variables at each step. GraphCast is notable for its strong deterministic forecasting performance versus the ECMWF Integrated Forecast System and for the careful ablations in the original paper.
Pangu-Weather
Huawei's Pangu-Weather uses a 3D Earth-specific Transformer with a hierarchical temporal stacking strategy: separate models trained for different lead times (1 hour, 3 hours, 6 hours, 24 hours) that can be composed to reach any horizon. Pangu was one of the first data-driven models to clearly beat operational NWP on standard benchmarks at specific lead times, and it kicked off the wave of papers that followed.
Head-to-head on forecasting accuracy
All three models are strong on the headline deterministic forecasting metrics — root mean squared error on 500 hPa geopotential and 2 m temperature being the most commonly reported. The ranking depends on lead time and variable.
- Short lead times (1-3 days). Pangu-Weather and GraphCast are typically strongest. Aurora is competitive.
- Medium lead times (3-7 days). GraphCast often edges out the others on deterministic benchmarks, though results vary by variable. Aurora catches up on several fields.
- Long lead times (7-10 days). All three models show degradation, and ensemble-based NWP still has meaningful advantages for probabilistic forecasting.
The practical takeaway: if you are building a pure global medium-range forecasting app, any of the three will work. Look at the variables and lead times your application cares about and pick accordingly.
Head-to-head on speed
The speed story is the part that sold the AI weather community on data-driven models in the first place. A physics-based operational NWP run takes hours on a supercomputer. A GraphCast, Pangu, or Aurora forecast takes seconds on a single GPU. That is a thousand-fold speedup, and it changes what you can do with weather data.
- You can run an ensemble of hundreds of forecasts interactively and get probabilistic output.
- You can condition forecasts on real-time observations as they come in.
- You can put a weather model inside an agent loop and have it respond in seconds.
None of those were practical before data-driven models. All three contenders unlock them. On raw inference speed the differences between the three are small compared to the gap between any of them and physics-based NWP.
Head-to-head on coverage
This is where the models separate. Forecasting is only one task an atmospheric model can do, and each model has taken a different approach to breadth.
GraphCast
Primarily a forecasting model. DeepMind has demonstrated strong results on deterministic forecasts and there is active research on ensemble and downstream uses. The core product is medium-range weather.
Pangu-Weather
Also primarily a forecasting model, with the hierarchical lead-time structure that allows composition to many horizons. Pangu's main contribution was demonstrating that a transformer-based data-driven model could beat operational NWP on a large benchmark set.
Aurora
The broadest coverage of the three. Aurora's foundation-model framing means the same backbone has been used for forecasting, air quality, and cyclone tracking, with more heads likely over time. If your application is wider than “predict the next 5 days of wind and temperature,” Aurora is the most natural starting point.
Which one to pick
The honest answer is: the one you can most easily access for your specific task.
- If you need medium-range deterministic forecasts and can run the weights locally, GraphCast is a strong default.
- If you need short-horizon forecasts at specific lead times, Pangu-Weather's hierarchical structure is a natural fit.
- If you need a single model for multiple atmospheric tasks or you want to avoid GPU management, Aurora through SciRouter's climate lab is the fastest on-ramp.
What 2026 looks like
The next twelve months are going to be interesting. Ensemble variants of all three models are being actively developed, which closes the probabilistic-forecasting gap with NWP. Fine-tuning heads for regional forecasting, nowcasting, and specialized variables like precipitation are under development. The foundation-model framing that Aurora introduced is likely to be adopted by the other research groups.
In the meantime, you can already do serious work with Aurora, GraphCast, or Pangu-Weather depending on your application. The choice is more about access and task fit than about which is objectively best.
Bottom line
Aurora, GraphCast, and Pangu-Weather are all meaningfully better than what was possible three years ago. Aurora wins on coverage and ease of access through SciRouter. GraphCast wins on deterministic medium-range forecasting accuracy. Pangu-Weather wins on hierarchical short-horizon use cases. None of them replace operational NWP, and all of them open applications that NWP speed could not support.