Sonny
StepNets one-step weather backbone: variable-aware embedding and a two-stage trunk, 69 ERA5 variables, 1.5° grid (120×240 lat×lon). Rows run from 88.5°S to 90°N (the south-pole row of the 121×240 WeatherBench 2 grid is dropped), and columns from 0° eastward.
Files
| File | Contents |
|---|---|
sonny_final.safetensors |
Sonny: EMA (decay 0.999) after a five-epoch learning-rate cooldown from epoch 50. These are the weights evaluated in the revised paper. |
model.safetensors |
Sonny (epoch 50): EMA (decay 0.9) at epoch 50, the weights of arXiv:2603.21284v1. |
config.json |
Constructor arguments. init_args.variables is the training order, which sets the order of the output channels. |
normalization/ |
normalize_mean, normalize_std, and normalize_diff_std_{6,12,24}, one value per variable. |
Correction (November 2026): earlier versions of config.json listed the first three variables as 2m_temperature, 10m_u_component_of_wind, 10m_v_component_of_wind. The training order is 10m_u_component_of_wind, 10m_v_component_of_wind, 2m_temperature. Outputs obtained with the earlier list have these three channels mislabeled. This applies to both weight files.
Usage
The model code is at github.com/jmj2316/Sonny and needs torch, timm, xformers, and a GPU. Run this from a clone of that repository:
import json
import numpy as np
import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from networks.sonny import Sonny
repo = snapshot_download("Francismj/sonny")
args = json.load(open(f"{repo}/config.json"))["init_args"]
variables = args["variables"]
model = Sonny(**args).cuda().eval()
model.load_state_dict(load_file(f"{repo}/sonny_final.safetensors"), strict=True)
def constant(name):
z = np.load(f"{repo}/normalization/{name}.npz")
values = np.concatenate([z[v] for v in variables]).astype("float32")
return torch.tensor(values).cuda().view(1, -1, 1, 1)
mean, std = constant("normalize_mean"), constant("normalize_std")
def step(x, hours):
"""x: physical units, (batch, 69, 120, 240), channels in the order of `variables`. hours: 6, 12, or 24."""
diff_std = constant(f"normalize_diff_std_{hours}")
interval = torch.full((x.shape[0],), hours / 10, device=x.device)
with torch.no_grad(), torch.autocast("cuda", dtype=torch.float16):
delta = model((x - mean) / std, variables, interval)
return x + delta.float() * diff_std
The scores in the paper average three rollouts, with steps of 6, 12, and 24 h, at every 24 h lead.
This procedure was run on the 355 daily 00 UTC starts of 2022 with sonny_final.safetensors. Over nine variables and ten lead days, the mean RMSE differs from the values in the paper by at most 0.47% in any cell, and by −0.13% on average. The paper's evaluation code scaled the increments in float16, which also left specific_humidity_{50,100,150,200} unchanged during rollouts; the code above scales them in float32.
Citation
If you use this checkpoint or architecture, cite:
Sonny: Breaking the Compute Wall in Medium-Range Weather Forecasting — Minjong Cheon, arXiv:2603.21284, 2026. https://arxiv.org/abs/2603.21284
- Downloads last month
- 88