WeatherWeatherLab

Open Weather Models Compared: GraphCast, Prithvi WxC, FourCastNet, Aurora, Pangu

Honest comparison of 5 open AI weather models: accuracy, lead time, licensing, commercial-use status, hardware needs. Which to pick for personal, commercial, and research use.

Ryan Bethencourt
April 19, 2026
14 min read

Five models, three honest questions

This comparison exists because most weather-model reviews bury the three questions that actually matter when you're building something real. (1) Is the license commercial? (2) What does it cost me per forecast? (3) What does it fail at? Everything else — architecture, parameter counts, benchmark charts — is secondary.

Below we compare five open AI weather models: GraphCast (DeepMind), Prithvi WxC (IBM/NASA), FourCastNet (NVIDIA), Aurora (Microsoft), and Pangu-Weather (Huawei). SciRouter hosts the three commercial-permitted models; the other two are covered in editorial context.

The honest license table

ModelLicenseCommercial use?On SciRouter?
GraphCast (1°)Apache 2.0✅ With attributionYes
GraphCast (0.25°)CC-BY-NC-SA❌ Research onlyNo (license)
Prithvi WxC 2.1Apache 2.0Yes
FourCastNet v2BSD-3-ClauseYes
AuroraMicrosoft Research❌ Research onlyNo (license)
Pangu-WeatherResearch/academic❌ Restricted redistributionNo (license)

GraphCast (Google DeepMind)

What it does: global medium-range forecast up to 10 days, trained on ERA5 reanalysis. Published in Science 2023, and in their paper DeepMind showed GraphCast beating the operational ECMWF HRES on ~90% of verification metrics out to 10 days — remarkable given HRES is the gold standard physical forecast.

Resolution: the public 1° variant on Hugging Face (~110 km per grid cell). The 0.25° variant (~25 km) is higher-fidelity but its weights are non-commercial.

Hardware: ~20 GB VRAM for 1°; ~40 GB for 0.25°.

When to pick GraphCast: 3–10 day planning where regional (not hyperlocal) resolution is acceptable. Pair with Prithvi WxC downscaling for kilometer resolution over a specific region.

python
# GraphCast 10-day forecast via SciRouter
import httpx
resp = httpx.post(
    "https://scirouter-gateway-production.up.railway.app/v1/weather/forecast",
    headers={"Authorization": f"Bearer {YOUR_API_KEY}"},
    json={
        "location": {"latitude": 35.0, "longitude": -97.0},
        "model": "graphcast",
        "lead_time_hours": 240,
    },
)

Prithvi WxC 2.1 (IBM + NASA)

What it does: atmospheric foundation model with 2.3B parameters. Its killer feature is downscaling — taking a coarse forecast (e.g., GraphCast 1° or NOAA GFS 0.25°) and producing a high-res regional grid.

Why it matters: the data-center-scale forecast models don't resolve sub-grid phenomena (cold-air drainage, urban heat islands, convection). Prithvi WxC brings the forecast down to 3 km over your region.

When to pick: you have a specific location where microclimate matters — agriculture, renewable energy siting, wildfire risk. Always used in combination with a base forecast model.

FourCastNet v2 (NVIDIA)

What it does: 0.25° global forecast, strong on extreme events (hurricanes, atmospheric rivers, heat waves). V2-small fits in 12 GB VRAM, making it the most deployable GPU-hosted option.

Benchmarks: In the v2 paper, NVIDIA reports Hurricane Ian track error within 50 km at +48h — comparable to NHC best-track. Strong short-range performance, weaker at 10+ days.

When to pick: short-range (1–6 day) forecasts where you want ML-class accuracy at higher resolution than GraphCast 1°. Especially good for extreme events.

Aurora (Microsoft Research)

What it does: 1.3B-parameter foundation model, claimed to match or beat GraphCast on many metrics. Can also predict air quality, ocean waves, and tropical cyclones.

Why it's not on SciRouter: the Microsoft Research License restricts use to research and non-commercial contexts. Hosting inference for paying users would violate § 2.3. We're watching for a commercial-permitted release.

What you can do with Aurora today: research, self-hosted academic work, or awaiting a relicense. Read the paper, run it locally, benchmark against your needs.

Pangu-Weather (Huawei)

What it does: one of the first pure-ML models to decisively beat operational NWP (published Nature 2023). 3D Earth-specific transformer, 0.25° resolution, strong on cyclone tracking.

Why it's not on SciRouter: the released weights are research/academic with restricted redistribution. Commercial hosting is ambiguous-to-prohibited.

Picking, one more time

  • Free baseline, any location: Open-Meteo (aggregates multiple sources). Zero credits.
  • US-only point forecasts + severe-weather alerts: NWS. 1 credit.
  • 3–10 day planning at 1° resolution: GraphCast. 5 credits.
  • Short-range (1–6 day), high-res, extreme events: FourCastNet. 5 credits.
  • Regional microclimate (agriculture, siting): Prithvi WxC downscaling. 10 credits.
  • Research context, no commercial constraint: also evaluate Aurora and Pangu locally.

Try them side-by-side

Open WeatherLab — the model selector shows license, cost, and latency per model. Run the same forecast through three different models and compare. 500 free credits/month.

Frequently Asked Questions

Which AI weather model is the most accurate in 2026?

For medium-range (3–10 day) global forecasts, GraphCast (Google DeepMind) and Aurora (Microsoft) trade the lead on published benchmarks; both beat operational NOAA GFS on ~90% of metrics. For short-range (1–3 day) at high resolution, FourCastNet v2 (NVIDIA) is strong on extreme events. For atmospheric downscaling, Prithvi WxC (IBM/NASA) is the state of the art. No single model wins every axis — picking is about use case, license, and compute.

Can I use GraphCast commercially?

Partially. The 1° resolution variant on Hugging Face is Apache 2.0 — free for commercial use with attribution. The higher-accuracy 0.25° variant is CC-BY-NC-SA — non-commercial only. This matters if you're building a paid product: host the 1°, reference the 0.25° only in editorial content. SciRouter hosts the 1° variant.

Why is Aurora excluded from SciRouter?

Microsoft Research's Aurora license restricts use to research and non-commercial purposes. Serving inference of Aurora to paying users would violate § 2.3 of the MSR License. We'd love to host it — we're watching for a commercial-permitted release. For now, read the paper and experiment locally.

Do these models need GPUs?

For inference, yes — but we handle that. GraphCast 1° needs ~20 GB VRAM, Prithvi WxC ~16 GB, FourCastNet v2-small ~12 GB. When you call the SciRouter API, you're hitting our RunPod-hosted A100 or A24 GPUs on pay-per-use billing (credits). No local GPU required.

What about Pangu-Weather?

Huawei's Pangu-Weather was one of the first ML weather models to convincingly beat operational NWP, and it's a benchmark reference in the literature. But the released weights are research-only with restricted redistribution. We chose not to host it. You can reference it in research; for commercial workflows, use GraphCast or FourCastNet.

How do these models compare to NOAA GFS?

GFS is a numerical (physics-based) model. GraphCast, FourCastNet, and Pangu are pure ML — they learn from historical reanalysis (ERA5) rather than solving physics equations. In published benchmarks, the ML models match or beat GFS on most variables beyond 24h lead time while running roughly 1000× faster at inference. GFS still has strengths: physical consistency, works anywhere on Earth at any scale, public domain.

Run this yourself — no GPU, no install

Free for researchers. Pick a tool, paste your input, see results in seconds.