WindBorne Systems has released WeatherMesh-6, the latest version of its AI-based weather forecasting model, claiming it can outperform leading physics-based and AI forecasting systems from the European Centre for Medium-Range Weather Forecasts, better known as ECMWF.
The startup’s biggest differentiator is not just AI. It is the proprietary atmospheric data collected by its own global balloon network.
TL;DR
- WindBorne says WeatherMesh-6 delivers more accurate medium-range forecasts than ECMWF’s IFS and AIFS ensembles.
- The model produces hourly forecasts and reaches 3 km resolution in Europe and the continental US.
- The company uses around 400 balloons in flight to feed real-time observations into its AI system.
Weather forecasting has traditionally been dominated by national agencies, supercomputers, and complex physics-based models. WindBorne Systems now wants to challenge that model with a combination of transformer-based artificial intelligence and its own atmospheric sensor network.
According to WindBorne, WeatherMesh-6 Global operates at 0.25-degree, or roughly 25 km, resolution and produced up to 38% lower ensemble-mean RMSE than ECMWF’s IFS, and up to 32% lower than ECMWF’s AIFS, across the evaluation window of July 2025 to March 2026. The company also said its 4.5-day forecast for 2-meter temperature was as accurate as a 1-day forecast from IFS.
TechCrunch reported that WeatherMesh-6 produces forecasts every hour, compared with the six-hour cycle used by many traditional models. It also noted that the model reaches 3 km resolution in Europe and the continental United States, where data quality is strongest.
“One simple way to understand it,” WindBorne chief product officer Kai Marshland told TechCrunch, is that WeatherMesh-6 is “as accurate five days out as a traditional forecast is the day before,” especially for surface temperature measurements.
The key to WindBorne’s pitch is data advantage. The company currently operates about 400 balloons in flight at any given time, launched from 15 sites worldwide. These balloons gather sensor readings that are fed into WindBorne’s forecasting stack.
WindBorne CEO John Dean framed this as the company’s core moat, saying, “I don’t understand, personally, the business model of being an AI based weather company without a dataset advantage.”
WindBorne says WeatherMesh is powered by transformer-based AI models running on GPUs, trained on decades of atmospheric data, and continuously updated by real-time balloon observations. The company also claims WeatherMesh can produce forecasts in seconds, up to 100,000 times faster than traditional numerical weather prediction systems.
The company has been building toward this moment for a while. In its WeatherMesh-3 research paper, WindBorne said WM-3 could generate 14-day global forecasts at 0.25-degree resolution in 12 seconds on a single RTX 4090, with up to 37.7% RMSE improvement over operational models in its evaluation.
Still, there is an important caveat. WindBorne’s WeatherMesh-6 performance numbers currently come largely from the company’s own disclosures and benchmark portal. WindBorne says it publishes raw gridded outputs for verification, but some model outputs are available only to researchers.
The startup’s broader mission is to close major weather data gaps. San Francisco Chronicle reported that WindBorne began in 2019 as a Stanford student project and has customers including NOAA and the Kenya Meteorological Department. The company’s long-term goal is to maintain 10,000 balloons aloft to observe the planet more comprehensively.
That ambition could reshape the weather intelligence market, especially for aviation, energy, agriculture, disaster planning, commodity trading, and public safety. However, WeatherMesh-6’s challenge to global forecasting giants will depend on whether its claimed gains hold up consistently in real-world, independently verified use.
For now, WindBorne’s message is clear: the future of weather forecasting may not be decided only by who has the biggest supercomputer, but also by who owns the most useful atmospheric data.
