GraphCast vs. GenCast vs. Pangu-Weather: AI Weather Models Compared

The short answer

All three models outperform traditional numerical weather prediction (NWP) on standard benchmarks, but they are built differently and suit different research needs. GraphCast is the well-established deterministic benchmark. GenCast is the right choice when you need probabilistic (ensemble) forecasts and uncertainty quantification. Pangu-Weather is fast and deterministic, with 3D atmospheric modeling that distinguishes it from the others architecturally.

Comparison table

	GraphCast	GenCast	Pangu-Weather
Developer	Google DeepMind	Google DeepMind	Huawei Cloud
Published	Science, 2023	Nature, 2024	Nature, 2023
Forecast type	Deterministic (single forecast)	Probabilistic (ensemble)	Deterministic (single forecast)
Architecture	Graph neural network	Diffusion model	3D transformer
Benchmark	Outperforms ECMWF HRES on 90% of targets	Outperforms ECMWF ENS on 97.2% of targets	Outperforms ECMWF HRES at 24h–168h
Uncertainty estimates	No	Yes — full ensemble distribution	No
Lead time	Up to 10 days	Up to 15 days	Up to 7 days
Open weights/code	Yes (Apache 2.0)	Yes	Yes (code + weights)
Inference speed	~1 min on TPU/GPU	Slower (ensemble generation)	Very fast — designed for speed

GraphCast — the established deterministic baseline

GraphCast (DeepMind, Science 2023) was the model that established AI weather forecasting as a serious alternative to traditional NWP. Using a graph neural network that represents the atmosphere as a set of connected nodes at different spatial locations and pressure levels, it produces a single deterministic global forecast up to 10 days ahead in roughly 1 minute on modern GPU/TPU hardware.

Use it when: You need a single best-estimate forecast and want the most-validated, best-understood open-source AI weather model. It’s a strong baseline for research comparing AI to NWP, and for applications that need fast forecast generation at high temporal cadence.

Key strength: Mature, well-documented, heavily benchmarked. More downstream research and tooling has been built around GraphCast than any other AI weather model.

Key limitation: Deterministic — gives you one forecast but no information about forecast uncertainty. If you need to know how confident to be in a prediction, this isn’t the right tool.

GenCast — probabilistic forecasting and uncertainty quantification

GenCast (DeepMind, Nature 2024) is the successor to GraphCast and addresses its core limitation: instead of a single forecast, it generates an ensemble — a set of many possible future weather states — using a diffusion model. The spread of that ensemble is your uncertainty estimate.

This matters because weather forecasting beyond a few days is inherently uncertain; a single predicted temperature value conceals that uncertainty rather than communicating it. Ensemble forecasts are the standard in operational meteorology for precisely this reason. GenCast brings that capability to AI models.

Use it when: You need uncertainty quantification — risk assessment, probabilistic climate projections, extreme event probability estimation, or decision-support applications where knowing the forecast spread is as important as the central estimate.

Key strength: The only one of these three that produces ensemble output. Beat ECMWF ENS (the operational probabilistic standard) on 97.2% of 1,320 verification targets.

Key limitation: Slower than GraphCast or Pangu-Weather because generating an ensemble requires running the diffusion model multiple times. Not the right choice if you need fast single forecasts.

Pangu-Weather — fast 3D transformer with strong short-range accuracy

Pangu-Weather (Huawei, Nature 2023) uses a 3D Earth Specific Transformer (3DEST) architecture that explicitly models the atmosphere in three spatial dimensions, rather than treating each pressure level separately. This 3D structure is architecturally different from GraphCast’s graph approach and may explain Pangu-Weather’s strong performance at shorter lead times (24–168 hours).

Use it when: You need very fast inference, care most about short-to-medium range accuracy (1–7 days), or want to compare AI model architectures for a research project.

Key strength: Extremely fast inference — milliseconds to seconds per forecast step on appropriate hardware. Strong documented performance at 24–72 hour lead times. The 3D transformer architecture is interesting from a modeling perspective and has influenced subsequent work.

Key limitation: 7-day effective lead time is shorter than GraphCast (10 days) or GenCast (15 days). Deterministic like GraphCast — no ensemble or uncertainty output. Less tooling and community support outside Huawei’s own platform.

Which one for your research?

Research question	Best tool
Benchmark AI vs. NWP for medium-range forecasting	GraphCast (most validated baseline)
Probabilistic prediction / uncertainty quantification	GenCast
Extreme event probability estimation	GenCast
High-throughput operational forecasting at speed	Pangu-Weather
Studying short-range (1–3 day) AI forecast accuracy	Pangu-Weather or GraphCast
Teaching or demonstrating AI weather forecasting	GraphCast (most documentation and examples)

What these models don’t do

All three are global medium-range atmosphere models, which means they:

Don’t produce convection-scale or local forecasts — they operate at grid resolutions of ~0.25° (~28 km), not the scales relevant to severe local storms
Don’t model ocean, land surface, or ice physics directly — they’re trained on ERA5 reanalysis data, not coupled earth-system models
Aren’t designed for climate projection — they extrapolate from current state; they’re not the right tool for century-scale scenarios
Require reanalysis data as input — typically ERA5 from ECMWF, which itself requires preprocessing pipelines

NVIDIA’s Earth-2 platform claims performance improvements across all three, but those comparisons are self-published rather than peer-reviewed. Treat them as unverified until independently benchmarked.