Which Weather Model Is Most Accurate? We Score Them Every Day, in the Open
FINAL-Bench
• • 7Darwin-Prediction-Weather
Darwin-Prediction-Weather is VIDRAFT's weather prediction system. It pairs a global forecast engine with a Darwin-27B decision core (ZTC readout on the RSI backbone).
- Forecast engine: produces the global numerical forecast. It records the lowest error in all 15 evaluated cells of the WeatherBench 2 deterministic protocol (2020), ahead of every model on the official scorecard.
- Decision core (the weights in this repository): Darwin-27B with the ZTC calibrated readout on the Darwin-27B-RSI backbone. Given a state and a typed question it returns a calibrated probability for every option in one forward pass. In this system it turns forecast fields into calibrated decisions: go or no-go, threshold exceedance, category probabilities.
Live leaderboard: verified every day against the world's weather centers
The Global Weather Model Leaderboard scores Darwin-Prediction-Weather every day against the operational models of ECMWF, NOAA, JMA, ECCC and MET Norway in 24 world cities. Each forecast is recorded when it is issued and scored when the observed value is published. Screens below were captured on 2026-11-13; the board updates daily.
1. No. 1 overall
7-day forecast of daily maximum temperature, 24 cities, 2026-02-01 to 2026-11-07, 4,848 verified forecasts. Darwin-Prediction-Weather records 2.09 °C RMSE, 13.2% lower error than No. 2 (ECMWF AIFS, 2.41 °C). Against the best physics model run on a supercomputer (ECMWF IFS, 2.84 °C) the lead is 26.4%.
2. Accuracy board
All eight models on one scale, scored on the cases every listed model covers.
| Rank | Model | Provider | Type | RMSE [°C] |
|---|---|---|---|---|
| 1 | Darwin-Prediction-Weather | VIDRAFT | AI | 2.09 |
| 2 | ECMWF AIFS | ECMWF | AI | 2.41 |
| 3 | AIGFS | NOAA NCEP | AI | 2.81 |
| 4 | ECMWF IFS | ECMWF | Physics (supercomputer) | 2.84 |
| 5 | MET Nordic / IFS | MET Norway | Physics (supercomputer) | 2.95 |
| 6 | JMA GSM | JMA | Physics (supercomputer) | 3.11 |
| 7 | GEM | ECCC | Physics (supercomputer) | 3.20 |
| 8 | GFS | NOAA NCEP | Physics (supercomputer) | 3.21 |
3. Lowest error at every forecast day, and the city map
Left: error grows with lead time for every model, and the Darwin-Prediction-Weather curve (dashed white) is the lowest from day 1 to day 7. Right: the most accurate model in each of the 24 cities.
4. No. 1 in every period and every lead time
Five periods (latest day, last 7 days, last 30 days, last 90 days, 2026 year to date) by seven lead times (1 to 7 days). Darwin-Prediction-Weather is No. 1 in all 35 cells.
City by city: 24 cities by seven lead times. Darwin-Prediction-Weather is No. 1 in 141 of the 168 cells for daily maximum temperature.
5. Monthly champion, every month on record
Darwin-Prediction-Weather holds every monthly title on the 2026 record.
6. Daily results and the 10-day outlook
For each city: the observed value and the winning model for each of the last seven days, followed by the Darwin-Prediction-Weather outlook for the next ten days. Darwin-Prediction-Weather has the most daily wins of the eight models.
7. Live Earth
The board's Live Earth view opens live global maps of wind, clouds, rain radar and temperature, and shows the 2026 champion of each city on a satellite map with live rain radar. Click a city on the board to see its forecast for the next days.
Leaderboard data: weather data by Open-Meteo (CC BY 4.0); verification with ERA5, Copernicus Climate Change Service. Live Earth imagery: NASA Global Imagery Browse Services (GIBS), part of NASA's Earth Science Data and Information System (ESDIS); rain radar by RainViewer.
8. No. 1 in every element
The board ranks four elements, each on its own tab. Darwin-Prediction-Weather is No. 1 in all four at every forecast lead. RMSE against ERA5, 24 cities, 2026 year to date (verified through 2026-11-07); the value in brackets is the next model, ECMWF AIFS in every cell.
| Element | 1 day | 3 days | 5 days | 7 days |
|---|---|---|---|---|
| Daily maximum temperature [°C] | 0.84 (1.35) | 1.17 (1.59) | 1.59 (1.90) | 2.09 (2.41) |
| Daily minimum temperature [°C] | 0.91 (1.50) | 1.09 (1.63) | 1.42 (1.85) | 1.81 (2.17) |
| Daily maximum wind speed [km/h] | 2.25 (3.60) | 2.70 (4.07) | 3.43 (4.76) | 4.39 (5.69) |
| Daily precipitation total [mm] | 4.15 (4.33) | 5.18 (5.24) | 6.80 (6.99) | 7.11 (7.30) |
For temperature and wind its error is 13 to 39 percent below the next model; for precipitation, 1 to 4 percent.
Each tab carries the same panels for its element: champion, accuracy board, error by forecast day, city map, daily results with the 10-day outlook, monthly champions and champions at a glance. The full daily record of every forecast and observed value is open in Weather-Model-Forecast-Archive.
WeatherBench 2 results (forecast engine)
RMSE vs ERA5, global, latitude-weighted, 1.5 degree, 2020, 702 initializations.
| Variable | 1 d | 3 d | 5 d | 7 d | 10 d |
|---|---|---|---|---|---|
| 2m temperature [K] | 0.468 | 0.743 | 1.107 | 1.531 | 1.984 |
| best on official scorecard | 0.488 | 0.786 | 1.125 | 1.552 | 2.022 |
| Z500 [m2/s2] | 33.7 | 108.5 | 240.6 | 400.3 | 595.4 |
| best on official scorecard | 36.8 | 113.8 | 250.7 | 415.1 | 606.5 |
| T850 [K] | 0.503 | 0.901 | 1.406 | 2.000 | 2.699 |
| best on official scorecard | 0.518 | 0.922 | 1.435 | 2.041 | 2.737 |
Official bests are WeatherMesh4 (1 to 7 days) and GenCast ensemble mean (10 days). Forecast files: VIDraft/Darwin-Prediction-Weather-WB2. Scorecard submission: google-research/weatherbench2#320.
Repository contents
| Path | What |
|---|---|
engine_config.json |
Forecast engine configuration: inputs, fusion strategies, and the per-variable, per-lead strategy table selected on 2018 |
model-*.safetensors, readout.safetensors, config.json, tokenizer |
Decision core weights, identical to Darwin-27B-ZTC-v9 |
ztc_server.py, autojev/ |
Serving and tooling for the decision core |
Forecast engine
For every initialization the engine produces one global forecast by gridpoint bias correction and skill-weighted fusion of its input forecasts. All correction fields and fusion weights use only data whose verification time precedes the initialization time. The strategy per variable and lead time was selected on 2018 data and frozen before the 2020 evaluation.
Inputs: WeatherBench 2 public forecast archives (GenCast ensemble mean, IFS ENS mean, GraphCast, Pangu-Weather, FuXi, IFS HRES). Ground truth: ERA5. Credit to the respective providers under their original licenses.
License
This repository is released under CC BY-NC 4.0. Non-commercial use only. For commercial licensing contact vidraft@mail.vidraft.net. Forecast dataset: CC BY-NC-SA 4.0.
VIDRAFT. Contact: vidraft@mail.vidraft.net
- Downloads last month
- 73
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for FINAL-Bench/Darwin-Prediction-Weather
Base model
FINAL-Bench/Darwin-27B-Opus Finetuned
FINAL-Bench/Darwin-27B-RSI