Darwin-Prediction-Weather: Korean AI Model Beats ECMWF, NOAA, and 6 Other Forecast Models Across 28 Benchmark Categories
TL;DR: VIDRAFT's AI weather forecasting engine, Darwin-Prediction-Weather, outperformed eight major numerical and AI weather prediction models — including ECMWF AIFS, NOAA AIGFS, and IFS — across all 28 evaluated forecast categories (4 variables × 7 lead times) in a 9-month, 24-city head-to-head study. For 7-day maximum temperature, Darwin achieved an RMSE of 2.08°C, 14.0% lower than the next-best model (ECMWF AIFS at 2.42°C). Developers working in climate tech, geospatial ML, or time-series forecasting should take note of this benchmark result.
What it is
Darwin-Prediction-Weather is a production-grade AI weather forecasting engine developed by VIDRAFT (비드래프트), a Korean Pre-AGI AI deep-tech startup. It is designed to generate global weather forecasts for the standard meteorological variables: maximum temperature, minimum temperature, wind speed, and precipitation, at lead times ranging from 1 to 7 days.
VIDRAFT conducted an independent accuracy study comparing Darwin against eight models from six major national and international meteorological agencies:
- ECMWF: AIFS (AI model) and IFS (physics-based numerical model)
- NOAA: AIGFS (AI model) and GFS (physics-based numerical model)
- DWD (Germany): ICON
- JMA (Japan): GSM
- Environment Canada: GEM
- MET Norway: MET Nordic
The evaluation covered 24 cities worldwide, ran for approximately 9 months (February 1 – November 11, 2026), and cross-referenced a total of 1.23 million forecast-observation pairs against real measured values. Darwin's forecasts were additionally submitted to Google Research's WeatherBench2, the standard open evaluation framework for global weather AI, using the same methodology as published entries on that leaderboard.
How it works
At a conceptual level, Darwin-Prediction-Weather is a data-driven AI model — not a physics-based numerical weather prediction (NWP) system. Rather than solving discretized atmospheric differential equations on a supercomputer grid (the approach of IFS, GFS, ICON, etc.), Darwin learns weather patterns directly from large-scale historical observational and reanalysis datasets.
This places it in the same architectural family as other modern AI weather models (e.g., ECMWF AIFS, Huawei Pangu-Weather, Google GraphCast), all of which use deep learning to approximate the mapping from an atmospheric state at time t to a predicted state at time t+n. The key differentiator, based on the benchmark results, appears to be in generalization quality: Darwin maintains accuracy advantages across diverse cities, seasons, and lead times rather than optimizing for a narrow evaluation regime.
No internal architectural details, training data sources, or hyperparameters have been publicly disclosed.
Benchmarks & results
All numbers below are sourced directly from the VIDRAFT/디지털타임스 study. The evaluation metric is RMSE (Root Mean Square Error); lower is better.
7-day maximum temperature forecast (4,896 matched cases):
| Model | RMSE (°C) | vs. Darwin |
|---|---|---|
| Darwin (VIDRAFT) | 2.08 | — |
| ECMWF AIFS | 2.42 | +14.0% |
| NOAA AIGFS | 2.83 | +36.1% |
| ECMWF IFS | 2.86 | +37.5% |
| MET Nordic | 2.94 | +41.3% |
| JMA GSM | 3.10 | +49.0% |
| NOAA GFS | 3.18 | +52.9% |
| GEM | 3.20 | +53.8% |
Additional headline results:
- 1-day max temperature: Darwin 0.84°C vs. AIFS 1.36°C — 38.6% smaller error
- 7-day min temperature: Darwin 1.81°C vs. AIFS 2.18°C — 16.7% smaller error
- 7-day wind speed: Darwin 4.36 km/h vs. AIFS 5.64 km/h — 22.9% smaller error
- Effective forecast horizon: Darwin's 7-day max temperature RMSE (2.08°C) is lower than AIFS's 6-day RMSE (2.14°C) and IFS's 5-day RMSE (2.21°C) — implying Darwin effectively sees 1 day further than ECMWF's AI model and 2 days further than ECMWF's physics model at equivalent accuracy
- Precipitation (next-day, ≥1mm binary detection accuracy): Darwin 90.6% vs. AIFS 88.4%
- Precipitation (7-day): Darwin 79.6% vs. AIFS 77.2%
- Darwin ranked #1 in all 28 categories (4 variables × 7 lead times), across all rolling time windows tested (last 7, 14, 30, and 90 days)
- City-level breakdown: Darwin ranked #1 in 7-day wind speed for all 24 cities, min temperature for 20 cities, max temperature for 18 cities
How to try it
Access to Darwin-Prediction-Weather is not publicly available at the time of this article. No Hugging Face model card, GitHub repository, or public API endpoint has been announced. VIDRAFT has submitted results to WeatherBench2 (Google Research's open weather AI leaderboard), so community-comparable scores may become visible there in the future. Watch VIDRAFT's official channels for any developer access announcements.
FAQ
Q: How does this compare to GraphCast or Pangu-Weather on WeatherBench2?
A: VIDRAFT submitted Darwin's hindcast results to WeatherBench2 using the standard evaluation methodology, which will enable direct comparison with GraphCast, Pangu-Weather, and other published models on that leaderboard. Specific WeatherBench2 scores were not included in the source article — check the WeatherBench2 public leaderboard for published entries.
Q: Is this an AI model or a traditional NWP system?
A: Darwin-Prediction-Weather is a pure AI/ML model, not a physics-based NWP solver. It is comparable in approach to ECMWF AIFS, NOAA AIGFS, and similar data-driven forecasting systems — all of which learn from historical atmospheric data rather than integrating physical equations on a numerical grid.
Q: Were the benchmark conditions fair across all models?
A: VIDRAFT applied identical evaluation conditions to all models: the same 24 cities, the same observation period, and — for the 7-day max temperature headline figure — the same 4,896 matched cases. The study ran for approximately 9 months of live daily forecasts compared against actual observations, reducing the risk of cherry-picked evaluation windows.
Originally reported by 디지털타임스 (2026-11-18) — source article.

