WeatherNext Memory Requirements and VRAM Usage: Full vs. Mini Models Compared

The full-resolution WeatherNext models require ≥80 GB VRAM (H100 GPU) and ≥25 GB system RAM, while the Mini models run on ≤16 GB VRAM (P100 GPU) and ≤12 GB RAM.

Memory and computational requirements vary dramatically between the two WeatherNext model families. Understanding these differences helps you select the right hardware configuration for your forecasting workload without encountering out-of-memory errors during inference.

Full-Resolution WeatherNext Models: Memory Footprint

The non-Mini WeatherNext models operate at full spatial resolution with complete pressure level coverage. According to the repository's README.md, these models demand high-end accelerators due to their large weight tensors and activation memory.

VRAM Requirements

The full models need at least 80 GB of GPU VRAM, targeting NVIDIA H100 GPUs as the reference hardware. Older GPUs—including 40 GB and 80 GB A100 variants—may struggle with the largest checkpoints. The repository explicitly states: "The non-Mini models require H100 for sufficient VRAM"【/cache/repos/github.com/google-deepmind/weathernext/main/README.md†L107-L108】.

System RAM Requirements

Host memory demands are equally substantial. For inference with comparable full-resolution models like GraphCast-small, you'll need at least 25 GB of system RAM when running on TPUv4 or A100 accelerators【/cache/repos/github.com/google-deepmind/weathernext/main/README.md†L223-L233】. This accommodates the input grids (ERA5/HRES data) and intermediate activations during autoregressive rollout.

WeatherNext Mini Models: Lightweight Alternative

The Mini models provide a deliberately constrained architecture for resource-constrained environments. These variants reduce spatial resolution to 1° and limit pressure levels, trimming both weight count and activation memory.

VRAM Requirements

Mini models fit comfortably within ≤16 GB of GPU VRAM, making them compatible with NVIDIA P100 GPUs and newer laptop-class accelerators. The repository confirms: "The Mini models should manage inference on a P100"【/cache/repos/github.com/google-deepmind/weathernext/main/README.md†L107-L108】.

System RAM Requirements

Host memory needs drop proportionally to ≤12 GB of system RAM. This enables deployment on standard workstations without specialized server hardware.

Loading and Running: Code Implementation

Both model tiers share identical APIs through weathernext.load_model(). The checkpoint name determines which variant loads and which memory constraints apply.

Loading a Full-Resolution Model

import weathernext as wn

# Requires H100/A100 GPU (≥80 GB VRAM)

full_model = wn.load_model(
    checkpoint="WeatherNextCyclones_2025",  # Full-resolution

    device="gpu",
)

full_forecast = full_model.autoregressive_rollout(
    initial_state=wn.utils.load_initial_state(...),
    num_steps=4,
)

Loading a Mini Model

# Runs on P100 or similar (≤16 GB VRAM)

mini_model = wn.load_model(
    checkpoint="WeatherNextCyclones_Mini_2024",  # 1° resolution

    device="gpu",
)

# Identical API, different memory footprint

mini_forecast = mini_model.autoregressive_rollout(
    initial_state=wn.utils.load_initial_state(...),
    num_steps=4,
)

The wn.load_model implementation in weathernext/weathernext1_cyclones/model.py handles weight-loading logic automatically. The same autoregressive_rollout function from weathernext/utils/autoregressive.py serves both model tiers—switching from full to Mini requires only a checkpoint name change【/cache/repos/github.com/google-deepmind/weathernext/main/weathernext/weathernext1_cyclones/model.py】【/cache/repos/github.com/google-deepmind/weathernext/main/weathernext/utils/autoregressive.py】.

Monitoring VRAM Usage

Verify memory constraints with JAX device utilities:

import jax
print(jax.devices()[0].memory_stats())  # Bytes allocated, peak usage

This helps confirm Mini models stay within the ~16 GB target during extended rollouts.

Hardware Recommendations by Use Case

Scenario	Recommended Model	Minimum GPU	Minimum RAM
Production forecasting, research	Full (non-Mini)	NVIDIA H100	25 GB
Prototyping, education, edge deployment	Mini	NVIDIA P100	12 GB
Colab demos, quick experiments	Mini	T4/V100 (free tier)	12 GB

The docs/weathernext2/wn2_demo.ipynb notebook demonstrates both model tiers and explicitly notes accelerator requirements for each【/cache/repos/github.com/google-deepmind/weathernext/main/docs/weathernext2/wn2_demo.ipynb】.

Summary

• Full models: ≥80 GB VRAM (H100), ≥25 GB RAM, full 0.25° resolution
• Mini models: ≤16 GB VRAM (P100), ≤12 GB RAM, reduced 1° resolution
• API parity: Same wn.load_model() and autoregressive_rollout() functions for both tiers
• Single-line switching: Change only the checkpoint name to toggle between model families

Frequently Asked Questions

Can I run full WeatherNext models on multiple smaller GPUs?

The repository documentation focuses on single-accelerator deployment. The weathernext.load_model() implementation does not expose multi-GPU sharding in the public API. For distributed inference, you would need to implement custom JAX parallelism around the core model weights in weathernext/weathernext1_cyclones/model.py.

Why do Mini models use 1° resolution instead of 0.25°?

The resolution reduction in Mini checkpoints directly decreases both weight tensor dimensions and activation grid sizes. This architectural constraint—documented in docs/weathernext1_cyclones/README.md—enables the 5× VRAM reduction while preserving core forecasting capabilities for demonstration and lightweight applications.

How do I confirm my GPU has enough memory before loading?

Check available device memory before model initialization using jax.devices()[0].memory_stats() or nvidia-smi. The Mini models target ≤16 GB peak allocation, but actual usage varies with batch size and rollout length. Start with short num_steps values and monitor memory growth during autoregressive_rollout() execution.

Are Mini model predictions significantly less accurate?

Mini models trade spatial precision for accessibility. The 1° resolution limits fine-scale feature representation compared to full 0.25° models, particularly for convective processes and localized extreme events. For synoptic-scale patterns and medium-range guidance, Mini outputs remain qualitatively useful—as demonstrated in the official Colab notebooks.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}