Choose models
Nine deterministic models publish through the same profile contract with different resolutions, lead times, pressure levels, optional capabilities, and field semantics. Two ensembles publish beside them, and Ensemble values covers their shape. Choose by lead time and required detail, then use disagreement between models to locate sensitivity in the day.
The slug in the first column is the string --model takes.
| Slug | Grid | New run every | Horizon | Note |
|---|---|---|---|---|
hrdps-west | 1 km | 12 h | 48 h | experimental feed |
hrdps-continental | 2.5 km | 6 h | 48 h | |
hrrr-conus | 3 km | 6 h | 48 h | |
nam-conus-nest | 3 km | 6 h | 60 h | retires 2026-10-06 |
rrfs | 3 km | 6 h | 84 h | experimental feed; NAM’s declared successor |
rdps | 10 km | 6 h | 84 h | |
nam | 12 km | 6 h | 84 h | retires 2026-10-06 |
gdps | 15 km | 12 h | 240 h | |
gfs | 25 km | 6 h | 384 h |
The catalogue (models.json, shipped in the package and read as the
model catalogue document) is the authority
behind every cell; per-model publication times and transfer costs are in the
feed reference.
Resolution versus horizon
Grid spacing and forecast reach trade differently across the catalogue; one number cannot stand in for model suitability.
HRDPS West: 1 kilometre grid and 48 hour horizon. HRDPS continental: 2.5 kilometre grid and 48 hour horizon. HRRR CONUS: 3 kilometre grid and 48 hour horizon. NAM CONUS nest: 3 kilometre grid and 60 hour horizon. RRFS: 3 kilometre grid and 84 hour horizon. RDPS: 10 kilometre grid and 84 hour horizon. NAM: 12 kilometre grid and 84 hour horizon. GDPS: 15 kilometre grid and 240 hour horizon. GFS: 25 kilometre grid and 384 hour horizon.
Choose by lead time and required terrain detail
| Question | Begin with | Then compare |
|---|---|---|
| What will the launch do today? | HRDPS 1 km (experimental) / 2.5 km, HRRR, RRFS (experimental) or NAM nest 3 km | each other, then observations |
| How stable is tomorrow’s forecast? | the same high-resolution models | RDPS |
| Is the weekend worth protecting? | RDPS, NAM 12 km, RRFS, GDPS | GFS trend |
| Is next week worth watching? | GDPS / GFS | wait for shorter-range guidance |
Both NAM entries carry a declared sunset in the catalogue (successor rrfs), so weigh that
date before building anything durable on them. The
feed reference records the lifecycle details, and
adopting a retiring model is the
case study behind the sunset declaration.
A global model can identify a ridge or trough many days out. It cannot resolve the launch cycle on a particular mountain face. Use long-range guidance to decide where to pay attention. Do not use it to choose an hour.
Grid resolution controls terrain and local detail
Each grid cell averages terrain and atmosphere across its footprint. A finer grid can place model terrain elevation closer to a launch and represent smaller weather features. A coarser grid describes the regional setup. Neither resolution guarantees the right answer.
Model terrain elevation changes pressure-level filtering and every height derived from the surface parcel. The Meteogram derivations define that dependency. Current grid spacing, domains, and verified model terrain elevations belong in the forecast model feed reference.
Time steps and pressure levels limit chart detail
Some Meteograms contain hourly columns; others contain three-hourly columns. Their curves may look equally continuous, but the three-hourly charts carry fewer observations of the model state. A narrow peak between three-hour steps is interpolation and does not add a forecast sample.
Vertical sampling imposes the same limit. Missing or widely spaced pressure levels reduce the detail available for lapse rate, wind shear, parcel crossings, and cloud layers. Compare the structure and sample positions behind a line before treating two smooth traces as equivalent evidence.
What each model declares
The discovery catalogue, rather than frontend assumptions, decides which fields and pressure levels exist.
HRDPS West (ECCC): publishes heat fluxes, 1 h max gust, CAPE, PBL height; 9 pressure levels from 925 to 600 hPa. HRDPS continental (ECCC): publishes heat fluxes, 1 h max gust, CAPE, PBL height; 14 pressure levels from 1015 to 600 hPa, with vertical velocity at 1000, 850, 700 hPa. HRRR CONUS (NOAA): publishes heat fluxes, instant gust, CAPE, CIN, PBL height, cloud layers; 9 pressure levels from 925 to 600 hPa, with vertical velocity at 925, 900, 875, 850, 800, 750, 700, 650, 600 hPa. RRFS (NOAA): publishes heat fluxes, instant gust, CAPE, CIN, PBL height, cloud layers; 9 pressure levels from 925 to 600 hPa, with vertical velocity at 925, 900, 875, 850, 800, 750, 700, 650, 600 hPa. RDPS (ECCC): publishes heat fluxes, 1 h max gust, CAPE, CIN, PBL height; 14 pressure levels from 1015 to 600 hPa, with vertical velocity at 850, 700 hPa. GDPS (ECCC): publishes heat fluxes, 1 h max gust, CAPE, CIN, PBL height; 14 pressure levels from 1015 to 600 hPa, with vertical velocity at 850, 700, 600 hPa. GFS (NOAA): publishes heat fluxes, instant gust, CAPE, CIN, PBL height, cloud layers, a cloud profile; 8 pressure levels from 925 to 600 hPa, with vertical velocity at 925, 900, 850, 800, 750, 700, 650, 600 hPa. NAM (NOAA): publishes heat fluxes, instant gust, CAPE, CIN, PBL height, cloud layers; 9 pressure levels from 925 to 600 hPa, with vertical velocity at 925, 900, 875, 850, 800, 750, 700, 650, 600 hPa. NAM CONUS nest (NOAA): publishes heat fluxes, instant gust, CAPE, CIN, PBL height, cloud layers; 9 pressure levels from 925 to 600 hPa, with vertical velocity at 925, 900, 875, 850, 800, 750, 700, 650, 600 hPa. REPS (ECCC): publishes heat fluxes; 5 pressure levels from 1000 to 500 hPa. GEPS (ECCC): publishes heat fluxes, CAPE, CIN; 5 pressure levels from 1000 to 500 hPa.
Generated from models.json; pressure columns share a 1015–500 hPa scale.
Use model disagreement to locate forecast sensitivity
Similar traces support the conclusion that the large-scale setup is straightforward. Separated traces identify sensitivity to timing, moisture, initialization, terrain, or model physics. A majority of models is not automatically correct. Related systems can share errors, and a coarse model can reach the right answer for the wrong local reason.
Use the comparison to locate the split.
| Pattern | First suspect |
|---|---|
| early timing split | boundary-layer development |
| persistent vertical offset | model terrain elevation or moisture |
| late-day fan | cloud development or collapsing surface heat |
The interactive timing comparison shifts development timing while holding its shape steady. Use the shared hour lens to expose the displacement, then inspect each model’s published profile.
The same development, shifted in time
Scrub one shared hour across two controlled profiles whose daytime development starts at different times.
An interactive comparison of two profiles with matching daytime-development shapes shifted to earlier and later hours.
Earlier development
Later development
Conclusion. Matching development shapes can occupy different hours; reading only one instant hides that timing displacement.