WeatherMesh-6
WeatherMesh-6 ( wm-6 ) is the next generation of WindBorne's AI weather model, initialized with our proprietary balloon observations from across the globe. WM-6 delivers improved accuracy through enhanced data assimilation and an ensemble built directly into the model, outputting calibrated percentiles and threshold probabilities, as well as raw ensemble members for a limited number of surface variables. WM-6 also introduces an expanded variable set and updates hourly, keeping forecasts current with the latest observations.
WM-6 also introduces a high-resolution regional model ( wm-6-3km ) producing 2.5 km regional surface forecasts over CONUS, updating every 15 minutes. Europe available soon.
1. Model Variants
WeatherMesh-6 is available in two variants:
| Variant | Code | Description |
|---|---|---|
| Global | wm-6 | Global 0.25° forecasts with 128 members processed to percentiles and threshold probabilities. |
| 3km | wm-6-3km | 2.5 km regional surface forecasts over CONUS. Europe available soon. |
2. Forecasting Regime
| Global | 3km | |
|---|---|---|
| Initialization Frequency | Hourly | Every 15 min |
| Forecast Horizon | 15 days | ~4 days |
| Ensemble Members | 128 | No |
| Data Latency | ~1 hour | ~15 minutes |
| Update Mode | Continuous | Continuous |
3. Data Assimilation
WM-6 uses WindBorne's enhanced AI-based data assimilation system, ingesting observations from multiple sources including WindBorne GSBs, geostationary satellites (GOES, SEVIRI), radar composites, and global surface networks (METAR, SYNOP) within a tight ±15-minute window around forecast zero. Like WM-5c, this independent assimilation enables continuous updates without dependence on external analysis cycles.
4. Model Resolution
| Global | 3km | |
|---|---|---|
| Models | wm-6 | wm-6-3km |
| Spatial Resolution | 0.25° (~25 km) | 2.5 km |
| Vertical Coverage | Surface variables + 25 pressure levels | Surface variables only |
| Available Domains | Global | CONUS Aligned to NOAA's URMA CONUS domain. Europe available soon. |
5. Outputs & Products
| Global | 3km | |
|---|---|---|
| Forecast Time Steps | 3-hourly | Hourly |
| Surface Variables | 39 | 8 |
| boundary_layer_height | ✓ | |
| cape | ✓ | |
| cloud_base_height | ✓ | |
| dewpoint_2m | ✓ | ✓ |
| direct_solar_radiation_3h | ✓ | |
| forecast_surface_roughness | ✓ | |
| high_cloud_cover | ✓ | |
| latent_heat_flux_3h | ✓ | |
| low_cloud_cover | ✓ | |
| max_temperature_2m_3h | ✓ | |
| mean_total_precipitation_rate | ✓ | |
| medium_cloud_cover | ✓ | |
| min_temperature_2m_3h | ✓ | |
| pressure_msl | ✓ | |
| runoff_3h | ✓ | |
| sensible_heat_flux_3h | ✓ | |
| short_wave_radiation | ✓ | |
| skin_temperature | ✓ | |
| snowfall_3h | ✓ | |
| soil_moisture_level_1 | ✓ | |
| soil_moisture_level_2 | ✓ | |
| soil_temperature_level_1 | ✓ | |
| soil_temperature_level_2 | ✓ | |
| solar_radiation_downwards_3h | ✓ | |
| surface_pressure | ✓ | |
| surface_solar_radiation_3h | ✓ | |
| temperature_2m | ✓ | ✓ |
| thermal_radiation_downwards_3h | ✓ | |
| top_solar_radiation_3h | ✓ | |
| total_cloud_cover | ✓ | |
| total_column_water_vapour | ✓ | |
| total_precipitation_3h | ✓ | |
| wind_gust_10m | ✓ | |
| wind_speed_10m | ✓ | |
| wind_speed_100m | ✓ | ✓ |
| wind_speed_250 | ✓ | |
| wind_u_10m | ✓ | ✓ |
| wind_u_100m | ✓ | ✓ |
| wind_v_10m | ✓ | ✓ |
| wind_v_100m | ✓ | ✓ |
| Upper-level Variables | 5 | — |
| geopotential | ✓ | |
| specific_humidity | ✓ | |
| temperature | ✓ | |
| wind_u | ✓ | |
| wind_v | ✓ | |
| Pressure Levels | 25 | — |
| 10 30 50 70 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 925 950 975 1000 | ||
| Delivery Format | Zarr or netCDF via cURL, PyPi package, or CLI | Zarr or netCDF via cURL, PyPi package, or CLI |
See Gridded Forecast API for usage.
6. Benchmarks
For a detailed look at WeatherMesh-6, see our introducing WM-6 blog post.
7. Historical Data
Historical data is not available via the API and is provided upon request. Please contact us for access.
| Global | 3km | |
|---|---|---|
| Available Period | April 2025 – March 2026 | April 2025 – March 2026 |
8. Known issues
- There are known artifacts in the historical data archive of WM-6 Global, affecting the following fields: total_precipitation_3h, runoff_3h, and mean_total_precipitation_rate. The artifacts manifest as "hot spots" with anomalously large, unphysical values. They are caused by an inference optimization deployed for the backtest. This issue does not affect operational forecasts of WM-6 Global. The artifacts can generally be removed in post-processing; please contact Haoxing Du at haoxing@windbornesystems.com if you need assistance.
- WeatherMesh-6 Global provides a deterministic forecast; however, we do not recommend using it in cases where forecast accuracy is important, and recommend that ensemble mean is used instead. Due to the model's architecture, all members are initialized from perturbed initial conditions, and the deterministic forecast is not a privileged, unperturbed forecast, resulting in a noticeable decrease in skill. The deterministic forecast is intended to convey spatial and correlational information, and is identical to the first member (member 0).