WeatherMesh transforms raw balloon observations into the world’s fastest, most accurate weather forecasts. By combining cutting-edge AI models with a unique global network of weather balloons, we deliver insights at a speed and scale that traditional physics-based forecasting cannot match.

Our record-breaking model, WeatherMesh

WeatherMesh (WM-4) is powered by transformer-based AI models running on GPUs, trained on decades of atmospheric data and continuously updated by real-time balloon observations. Unlike traditional Numerical Weather Prediction (NWP), which requires vast supercomputing resources and hours of runtime,

WeatherMesh produces forecasts in seconds—up to 100,000x faster and with a fraction of the compute cost.
View API Docs

Forecast Accuracy & Impact

WeatherMesh is more accurate than ECMWF's HRES, the gold-standard physics-based global forecasting system, across all variables and lead times from 1-10 days. The accuracy improvement is particularly notable for surface variables like 2-meter temperature and in the 7-10 day range.

View Live Benchmarks

Powered by Our Real-Time Balloon Data

WeatherMesh is uniquely informed by our global balloon network. AI-based data assimilation means forecasts refresh every 10 minutes, providing the most up-to-date picture of evolving weather systems.

Why WeatherMesh is Different

Better Accuracy

More accurate at predicting temperature at ground weather stations than leading AI models such as AIFS

Higher Temporal Resolution

WeatherMesh is able to forecast in hourly and 3-hourly timesteps (vs. 12-hourly from other models like GenCast), enabling more actionable insights

Open Science

We publish at conferences, maintain realtime benchmarks, and open-source our code, ensuring transparency and advancing the field—unlike closed competitors

Technical details

WM-4

Transformer-based AI model

State-of-the-art neural network architecture optimized for processing weather data

Modular architecture

Encoder-processor-decoder structure enables flexible training and efficient inference

Latent space representation

Weather states are mapped into latent space representations for efficient pattern recognition

Comprehensive dataset

Trained on 50+ years of historical weather data

For a complete technical deep-dive into WM-4's architecture, training methodology, and performance benchmarks, read our detailed technical blog post.