NOAA GFS

NOAA's global model, in the public domain: four runs a day on a 13 km grid, wind hourly to 5 days then every 3 hours out to 16 days.

Now in Saily WR

Model Type Status Run Frequency Resolution Horizon
United StatesGFSWindUp to date29/10 06z4×/day0.25°15 d 14 h

Seventy years of American numerical forecasting

The GFS descends from the very first operational numerical forecasts. In 1954 the US Weather Bureau, Air Force and Navy founded the Joint Numerical Weather Prediction Unit; it installed an IBM 701 in March 1955 and chose Jule Charney's three-level Princeton model, but only from 1958, with a simpler model on a more powerful IBM 704, did its forecasts become useful. In January 1958 it merged with the National Weather Analysis Center to form the National Meteorological Center, which became the National Centers for Environmental Prediction (NCEP) in October 1995.

The direct ancestor of today's model is the global spectral model that went into service in August 1980, at a resolution of about 375 km with 12 levels; spectral means the fields were described as sums of waves around the globe rather than values on a grid. For two decades it ran under two names: the AVN, run four times a day out to about five days, and the MRF, the once-a-day medium-range forecast to about two weeks. In 2002 the name Global Forecast System replaced both, for what had become a single run four times a day out to 16 days.

The spectral GFS grew finer step by step, from about 55 km in 2002 to a semi-Lagrangian version at T1534, about 13 km, in January 2015; in 2007 its analysis moved to the GSI code still in use today. In July 2016 NOAA chose a new dynamical core, the heart of the model that computes wind and pressure: FV3, developed at its Geophysical Fluid Dynamics Laboratory in Princeton, keeping the GFS name. GFS v15, the first FV3 version, became operational on 12 June 2019, the first change of the model's dynamical core in almost 40 years.

GFS v16, operational since 22 March 2021, doubled the number of vertical layers from 64 to 127, raised the model top from 54 to 80 km, renewed the turbulence scheme and brought the global wave model inside the GFS. It is still the operational version in October 2026, as 16.3.x maintenance releases. Its successor, GFS v17, was proposed in April 2026 for October 2026: about 9 km, coupled to an ocean, a sea-ice and a wave model, with new assimilation built on the JEDI framework; as of 22 October 2026 NOAA had not yet announced its implementation date.

The model: FV3 on a cubed sphere

FV3 stands for finite-volume cubed-sphere. Picture a cube inflated until its faces cover the globe: each of the six faces is cut into a grid, and in the C768 grid of the GFS each face is 768 cells by 768, about 13 km apart. This avoids the crowding of meridians at the poles that troubles latitude-longitude grids. Finite-volume means the model tracks how much air, heat and moisture flows across the walls of each cell, so that mass is conserved exactly.

The core is non-hydrostatic: it does not assume that vertical motion is negligible, which matters for strong convection. In the vertical it uses 127 layers from the surface up to about 80 km, packed most tightly near the ground where the wind a sailor feels is shaped. Because the GFS is global, there are no lateral boundaries: what leaves one side of the grid simply re-enters on another face.

Everything smaller than a cell is parameterised, that is represented by simplified rules fed by the resolved fields. Showers and thunderstorms come from a mass-flux convection scheme of the Arakawa-Schubert family; clouds and rain from the GFDL microphysics, which carries several kinds of cloud water and ice; turbulence near the surface from a scale-aware scheme based on turbulent kinetic energy and eddy-diffusivity mass-flux, introduced with v16. Radiation uses the RRTMG code, the land the Noah land-surface model, and a near-sea-surface temperature scheme lets the top of the ocean warm by day and cool by night.

At 13 km the GFS resolves fronts, lows and the large features of the trade winds and monsoons, but not individual squalls, sea breezes in narrow bays or the funnelling of wind between islands and through straits. Mountains are smoothed to the grid, so a 2,000 m ridge on a small island may appear as a low hump. The wave model WAVEWATCH III runs inside the GFS, driven by its winds without feeding back to the atmosphere.

From observations to GRIB files

Each run starts from an analysis made by the Global Data Assimilation System (GDAS), which blends a short GFS forecast with observations from satellites, radiosondes, aircraft, ships, buoys and land stations. Its method, called hybrid 4D-EnVar, finds the state of the atmosphere that best fits both the observations, spread over a time window, and the background forecast; the errors it expects of that forecast come partly from a fixed statistic and mostly from an 80-member ensemble of short forecasts at about 25 km. The computation is done by NCEP's Gridpoint Statistical Interpolation code, GSI.

Speed and completeness pull against each other, so NCEP runs two analyses per cycle. The GFS forecast starts from an early analysis, made soon after the cycle time so that the forecast reaches users quickly; the GDAS waits longer for late observations and its more complete analysis prepares the next cycle. Since v16, the analysis increment is spread gradually over the first hours of the forecast rather than added all at once, which avoids a start-up shock.

The runs at 00, 06, 12 and 18 UTC are computed on WCOSS2, NOAA's twin HPE Cray supercomputers Cactus and Dogwood, in Phoenix, Arizona and Manassas, Virginia, operational since June 2022 at 12.1 petaflops each; one runs production while the other stands by. Each run forecasts 16 days, written hourly to 120 hours and every 3 hours beyond. In July 2026 NOAA announced that it would move these operations, the GFS included, to Google Cloud from early 2027, with the move due to end by December 2027.

The output is published in GRIB2 on regular grids, the finest at 0.25°, on NOMADS, the NOAA Operational Model Archive and Distribution System, which also lets users cut out a region and a few variables. As information produced by the US government it is in the public domain: free of charge, for any lawful purpose, including commercial, provided it is not presented as official NOAA material once modified.

The GFS at sea

Being free and unrestricted for decades has made the GFS the default GRIB of many offshore sailors, often the first file downloaded over a satellite link. Its hourly steps to five days and its 16-day reach cover a whole ocean passage, and for the large picture, the position of the Azores High, the track of a depression across the Atlantic, the extent of the trades, it is a sound reference.

Its limits are those of a 13 km global model. Near the coast it misses sea breezes, the acceleration of the wind around headlands and its funnelling in straits, and the calm in the lee of a high island; squalls under a tropical cloud line are averaged into a steady wind. In a tropical cyclone it gives a useful track, but a 13 km grid cannot hold the tight core of a strong hurricane, so the peak wind is usually too low.

NCEP's own evaluation of v15 found that it tended to move weather systems too fast and was less skilful on the tracks of strong tropical cyclones, two faults v16 reduced. Since the GFS starts from its own analysis, made with different methods from ECMWF's, its errors are largely independent of those of the IFS and AIFS, which share an analysis. That makes it the natural first comparison: when the GFS and the IFS agree on a route, the forecast is probably solid; when they differ, the timing of a front or the position of a low is uncertain, and the choice of route should leave room for both.

GFS in Saily WR

Saily WR downloads each new GFS run from NOMADS as soon as it is published, four times a day, on the 0.25° grid: the 10 m wind, as its east-west and north-south components, hourly to 120 hours and then every 3 hours to 384 hours, 16 days ahead. With NOAA AIGFS, it is the model with the longest horizon in the app, and because it is global, every route gets a GFS version.

The isochrone route is sailed from your boat's polar in that wind, interpolated in space between grid points and in time between steps, and is shown next to the routes of every other model that covers both your start and your finish. The sea current and land avoidance you choose apply to it exactly as to the others. The Data page shows live which GFS run the app is on and how far ahead it reaches.

Beyond ten days or so, a GFS route says more about the general pattern than about the exact track. Read its last days as a sketch, compare it with the IFS and AIFS routes, which stop at 15 days, and recompute as new runs arrive.

Sources

  1. Global Forecast System (GFS), NOAA/NCEP Environmental Modeling Center (2026)
  2. U.S. supercomputers for weather and climate forecasts get major bump, NOAA (2022)
  3. NCEP global models history table, NOAA/NCEP Environmental Modeling Center (2026)
  4. Our History, NOAA/NCEP Environmental Modeling Center (2026)
  5. NOAA to develop new global weather model, NOAA (2016)
  6. Global Forecast System FV3 fact sheet, National Weather Service (2019)
  7. NOAA's use of cloud infrastructure grows to include weather prediction models, NOAA (2026)
  8. GFSv16 webinar, F. Yang, NOAA/NCEP EMC (2020)
  9. Public Information Statement 26-29: proposed upgrade to the GFS (v17), National Weather Service (2026)
  10. Global Ensemble Forecast System (GEFS), NOAA/NCEP Environmental Modeling Center (2026)
  11. NOAA Operational Model Archive and Distribution System (NOMADS), NOAA/NCEP (2026)
  12. Disclaimer, National Weather Service (2026)

All the data routes are computed on