GenCast
DeepMind's diffusion-based ensemble weather forecasting model — produces probabilistic forecasts that beat ECMWF's operational ensemble system on 97.2% of verification targets, with open weights and code.
What it does
GenCast is an AI weather forecasting model from Google DeepMind that uses a diffusion-based approach to generate ensemble forecasts — sets of many plausible future weather states rather than a single deterministic prediction. This matters because a single forecast can’t tell you how confident to be in the prediction; an ensemble can. Published in Nature (December 2024), GenCast is the follow-up to the earlier GraphCast model and addresses GraphCast’s main limitation: GraphCast produces one deterministic forecast, whereas GenCast produces a distribution over forecasts that quantifies uncertainty.
The peer-reviewed benchmark numbers are strong: GenCast beat ECMWF’s ENS (the operational ensemble system that professional meteorology has used as the gold standard) on 97.2% of 1,320 verification targets, with that win rate climbing to 99.8% at lead times beyond 36 hours. Code, weights, and sample forecasts are publicly released.
Best for
Research requiring uncertainty quantification in weather or climate forecasting — risk assessment, decision-support applications, probabilistic climate projections, or any work where a single forecast is insufficient. The ensemble approach also makes it useful for studying predictability: how forecast uncertainty grows with lead time, and where the atmosphere is inherently less predictable.
Pricing
Free and open-source. Code, weights, and sample forecasts released by DeepMind. Requires GPU compute to run; no hosted inference API.
Strengths
- Peer-reviewed Nature publication — the benchmark numbers are independently verifiable, not vendor-published
- Generates probabilistic forecasts where GraphCast produces only a single prediction — the right tool when you need uncertainty estimates
- Outperforms ECMWF ENS, the long-standing professional meteorology ensemble benchmark
- Open weights and code allow for research use, fine-tuning, and adaptation
Limitations
- Requires GPU hardware and ML infrastructure to run — not a point-and-click tool
- Like all AI weather models, performs less reliably on extreme events (severe convection, extreme precipitation) outside its training distribution
- Does not incorporate physical constraints; its ensemble spread is statistically learned rather than physically derived, which matters for mechanistic interpretation
- Comparison benchmark (ENS) is from ECMWF; ECMWF’s own AIFS model is also now operational and may be a more current comparison point for some use cases
- NVIDIA’s competing Earth-2 suite claims to outperform GenCast across more than 70 variables, but that comparison has not yet been independently peer-reviewed — treat it as unverified until published
How it compares
| vs. | Key difference |
|---|---|
| GraphCast | GraphCast is deterministic (one forecast); GenCast is probabilistic (an ensemble). Use GenCast when you need uncertainty quantification |
| ECMWF AIFS | AIFS is operational at a major agency, not just a research release; GenCast has stronger published benchmark numbers but AIFS is what operational meteorology currently runs |
| NVIDIA Earth-2 | Earth-2 claims better performance but comparison is self-published, not peer-reviewed; GenCast’s Nature paper makes it the current benchmark with independently verifiable numbers |
| Pangu-Weather | Pangu-Weather is deterministic like GraphCast; GenCast’s probabilistic approach serves a different use case |
Related content
- Field Guide: Climate & Earth Science
- News: GenCast’s Real Numbers, and a New Open-Source Challenger
- Glossary: Ensemble Forecasting, Reanalysis Data