GFS vs HRRR vs NAM: Which Weather Model Should You Trust?

When you check your weather forecast, you're seeing the output of powerful computer models that simulate the atmosphere's behavior. But which model is best? Weather enthusiasts know that different models often disagree, sometimes significantly. Understanding the strengths and weaknesses of major forecast models helps you interpret weather predictions more intelligently, whether you're a storm chaser, pilot, farmer, or just someone who wants to know if they need an umbrella.

Major change coming October 6, 2026. NCEP is retiring the NAM, SREF, HREF, HiresW, and NAM MOS at 12 UTC that day, replacing them with the Rapid Refresh Forecast System (RRFS) and its ensemble, REFS. RRFS is a 3-km North America model updated hourly out to 18 hours, with runs out to 84 hours four times a day; REFS is scheduled to take over the ensemble roles of HREF and SREF. RRFS is covered first below as the regional model to learn, and the NAM section is kept for reference. The change can be postponed for significant weather, so check the current service notices before treating NAM and SREF as retired. Details are in Service Change Notice 26-47 and SCN 26-48, and current notices are indexed at weather.gov/notification.

Why Models Disagree

Before diving into specific models, it helps to understand why forecasts differ.

Weather models are numerical simulations that divide the atmosphere into three-dimensional grid boxes. They calculate temperature, pressure, humidity, and wind at millions of points, stepping forward in time using equations that govern atmospheric physics.

Different models make different choices about:

Resolution. Higher-resolution models use smaller grid boxes and can capture smaller-scale features like thunderstorms. Lower-resolution models smooth out these details but may handle large-scale patterns better.

Physics parameterizations. How do you simulate cloud formation, radiation, soil moisture, and turbulence? Models use different mathematical approximations for processes too small to explicitly resolve.

Domain. Some models cover the entire globe; others focus on specific regions like North America. Regional models can afford higher resolution because they're not simulating the whole planet.

Update frequency. Some models run every hour; others run every 6 or 12 hours. More frequent runs can capture rapidly evolving situations.

Data assimilation. Models ingest observational data (from satellites, weather stations, radiosondes, aircraft) to initialize their simulations. Different models use different data sources and assimilation techniques.

No model is perfect because the atmosphere is chaotic, tiny unmeasured perturbations can amplify into significant forecast differences over time. That's why meteorologists examine multiple models and use human judgment to create their forecasts.

Enlarge

The GFS: Global Workhorse

Full name: Global Forecast System Operated by: NOAA's National Centers for Environmental Prediction (NCEP) Resolution: 13 km (about 8 miles) Domain: Global Run times: Every 6 hours (00z, 06z, 12z, 18z) Forecast range: Out to 16 days

The GFS is America's primary global weather model and one of the most widely used models in the world. It's free, frequently updated, and provides forecasts for the entire planet out to more than two weeks.

Strengths:

• Long-range forecasting out to 16 days
• Frequent updates (four times daily)
• Handles large-scale weather patterns well
• Free and widely available
• Good for tracking major systems like hurricanes and nor'easters

Weaknesses:

• Lower resolution misses small-scale features
• Can struggle with exact positioning of precipitation boundaries
• Less accurate for localized events like afternoon thunderstorms

Best for: Big-picture outlook, long-range planning, tracking major storm systems, hurricane forecasting

The European Model (ECMWF)

Model: Integrated Forecasting System (IFS), medium-range control forecast Operated by: European Centre for Medium-Range Weather Forecasts (ECMWF) Resolution: 9 km (about 5.5 miles) Domain: Global Run times: Every 6 hours (00z, 06z, 12z, 18z) Forecast range: Out to 15 days from 00z and 12z; the 06z and 18z runs stop at 6 days

The IFS is commonly called "the Euro." It provides global guidance for medium-range forecasting. Performance comparisons depend on the variable, region, season, and forecast lead time; a model ranking is not a guarantee for a particular storm.

Strengths:

• Global medium-range guidance
• Excellent for medium-range forecasting (days 3-10)
• Guidance for comparing possible hurricane tracks
• Higher resolution than GFS
• Strong ensemble forecasting system

Weaknesses:

• Free open-data products are available, but some delivery services and third-party tools charge fees
• Runs less frequently than some models
• Still limited resolution for mesoscale events

Best for: Medium-range forecasting, high-stakes weather decisions, hurricane tracking, professional meteorology

ECMWF provides free open forecast data. Its public subset and paid delivery options differ in available products and access methods; do not assume all ECMWF forecasts require payment.

RRFS and REFS: The Scheduled Regional Transition

Full names: Rapid Refresh Forecast System, and RRFS Ensemble Forecast System Operated by: NCEP Resolution: 3 km, convection allowing Domain: North America, with a CONUS focus Run times: RRFS hourly, with longer runs four times a day; REFS four times a day Forecast range: RRFS out to 18 hours on the hourly cycles and 84 hours on the 00, 06, 12 and 18 UTC cycles; REFS out to 60 hours Scheduled operational start: October 6, 2026, subject to the NWS service notice

RRFS is the regional model to learn. At implementation, RRFS will take over roles served by NAM and HiresW, while REFS will replace HREF and SREF ensemble products. The HRRR is not part of that consolidation: it keeps running, and it supplies two of the REFS members over the CONUS and Alaska domains. For severe weather that probabilistic guidance is the part forecasters actually lean on.

Strengths:

• Convection-allowing at 3 km, allowing the model to represent storm-scale features while still parameterizing unresolved processes
• Hourly updates, which matters most on the day of an event
• One system in place of four, so there are fewer separate products to track. Disagreement still has to be interpreted: REFS members carry differing physics by design, and two of them come from the HRRR
• REFS gives probabilities rather than a single deterministic run

Weaknesses:

• Operational biases are not established yet. Comparisons from the parallel runs are provisional, and it takes seasons of verification before anyone can say confidently how a model errs
• Consolidation means less independent guidance to cross-check against
• Domain limited to North America, like the NAM before it

Best for: Days 1 to 3 forecasting over North America, severe weather environments, and precipitation detail. If your workflow currently depends on NAM soundings, HREF probability fields, or SREF plumes, look at the RRFS and REFS output now so you learn its behavior before you need it during an event.

The NAM: What RRFS Will Replace

Retiring at 12 UTC on October 6, 2026. NAM remains operational ahead of the scheduled transition. RRFS parallel output is available for evaluation; use the operational products until NWS confirms the changeover.

Full name: North American Mesoscale Model Operated by: NCEP Resolution: 12 km parent, with a 3 km (about 2 miles) CONUS nest Domain: North America Run times: Every 6 hours Forecast range: Parent out to 84 hours (3.5 days); the 3 km nest out to 60 hours

The NAM focuses on North America at higher resolution than global models. Its nested 3-km CONUS domain can resolve individual thunderstorm cells, which makes it useful for severe weather forecasting.

Strengths:

• Its nest represents smaller-scale features than the parent grid
• Severe weather and convection forecasting
• Precipitation timing and amounts for smaller-scale events

Limitations:

• Limited forecast range (84 hours on the parent, 60 on the 3 km nest)
• Domain limited to North America
• Less reliable in later forecast hours

NAM MOS is being discontinued at the same time, alongside SREF, HREF, and HiresW.

The HRRR: Rapid Refresh

Full name: High-Resolution Rapid Refresh Operated by: NCEP Resolution: 3 km (about 2 miles) Domain: Separate CONUS and Alaska domains CONUS run times: Every hour CONUS forecast range: Out to 18 hours on standard cycles, 48 hours on the 00, 06, 12 and 18 UTC cycles

The HRRR is the go-to model for nowcasting and short-range forecasting. Its hourly updates and high resolution make it invaluable for tracking active weather situations.

Strengths:

• Hourly updates capture rapidly evolving weather
• 3-km resolution resolves individual thunderstorms
• Excellent for severe weather timing
• Short-range guidance relevant to aviation weather
• Captures terrain-driven phenomena
• Assimilates radar data

Weaknesses:

• Very short forecast range (48 hours)
• Regional coverage, with separate CONUS and Alaska domains
• Can sometimes spin up spurious convection
• Computationally intensive

Best for: Day-of forecasting, storm chasing, aviation, timing of precipitation, active weather situations

When storms are developing and you need to know where they'll be in the next few hours, the HRRR is your best friend. Storm chasers rely heavily on HRRR guidance.

The RAP: HRRR's Parent

Full name: Rapid Refresh Operated by: NCEP Resolution: 13 km Domain: North America Run times: Every hour Forecast range: 21 hours on standard cycles; 51 hours on the 03, 09, 15 and 21 UTC cycles

The RAP is the coarser-resolution parent of the HRRR. It covers a larger domain (all of North America) while still providing hourly updates.

Strengths:

• Hourly updates
• Covers all of North America including Alaska
• Good for aviation weather across the continent
• Provides boundary conditions for HRRR

Weaknesses:

• Lower resolution than HRRR
• Can't resolve individual thunderstorms
• Short forecast range

Best for: Hourly updates for areas outside CONUS, aviation weather for Alaska and Canada

Canadian Models

Full names: Canadian Global (GDPS), Canadian Regional (RDPS), Canadian High-Resolution (HRDPS) Operated by: Environment and Climate Change Canada

Canada runs its own suite of models that are particularly valuable for weather affecting the northern US and Canada.

Strengths:

• Excellent for weather systems affecting Canada
• Good Arctic coverage
• HRDPS provides high-resolution detail for populated Canadian regions

Limitations:

• Domain, resolution, and forecast range differ across GDPS, RDPS, and HRDPS; check the specific product

Best for: Weather events affecting Canada and the northern tier of the US

Ensemble Models: Embracing Uncertainty

Individual model runs show one possible evolution of the atmosphere. Ensembles produce a range of possible outcomes by varying initial conditions and, depending on the system, boundary conditions, model physics, or the models themselves. During hurricane season those members are what fill a spaghetti model plot.

GFS Ensemble (GEFS): 31 members, shows the range of possible GFS solutions

ECMWF Ensemble (EPS): 51 members, highly regarded for probabilistic forecasting

SREF: Short-Range Ensemble Forecast, multiple models combined. Retiring October 6, 2026, replaced by REFS

REFS: RRFS Ensemble Forecast System, the convection-allowing short-range ensemble that takes over from SREF and HREF

Ensembles help quantify uncertainty. Close agreement can support confidence, but members can share errors and miss plausible outcomes. If they show widely divergent solutions, the forecast is uncertain. The spread of ensemble members is as important as the mean.

How to Use Models Wisely

Understanding model strengths helps you interpret forecasts more intelligently:

For events more than 5 days away: Focus on the Euro and GFS. Watch for agreement or disagreement between them. Check ensemble spreads.

For events 2-5 days away: Compare the regional guidance against GFS and Euro. Use the NAM until NWS confirms the scheduled RRFS transition; neither regional model covers the full five days, and RRFS output is available now as a pre-implementation parallel feed on NOMADS. The higher resolution adds value for regional details, but verify against global models for larger-scale patterns.

For tomorrow and the next day: The regional convection-allowing models become primary tools. Frequent updates help refine timing as the event approaches.

For active weather (today): The HRRR is the workhorse. Check it hourly during severe weather events, and compare radar observations to its predictions to calibrate confidence. Note that verification varies by season, region, and parameter, so "best" is situational rather than absolute.

For storm chasing: Start with Day 1 SPC outlooks and model soundings for mesoscale environment analysis, from RRFS and GFS once RRFS is operational, scheduled for October 6, 2026, and from the NAM and GFS until then. On the chase day, live on the HRRR, with REFS taking over the ensemble role from HREF and SREF at the same changeover.

For aviation: The RAP and HRRR provide critical aviation weather guidance. NAM MOS is discontinued alongside the NAM on October 6, 2026.

Model Biases to Watch

Biases depend on model version, region, season, variable, and forecast lead time. Avoid treating a model as permanently "too fast," "too wet," or "too cold." Those shortcuts can fail when the weather pattern or model configuration changes.

Compare recent forecasts with observations for the conditions you care about. Read official forecast discussions for the current event, and compare ensemble outcomes. If one model predicts arrival at 6 AM and another at 10 AM, the correct time is not necessarily halfway between them.

For RRFS and REFS, evaluate the parallel output while keeping its pre-operational status clear. A replacement model does not automatically inherit the strengths or biases of the products it replaces.

The Future of Forecasting

Weather modeling continues to advance rapidly:

Higher resolution: Models continue to increase resolution as computing power grows. The HRRR's 3-km grid was science fiction not long ago.

Better physics: Improved parameterizations of cloud processes, boundary layer turbulence, and surface interactions make models more accurate.

Machine learning: AI techniques are augmenting traditional numerical weather prediction, particularly for post-processing model output.

Ensemble techniques: Larger ensembles and better methods for extracting probabilistic information improve uncertainty quantification.

Data assimilation: New observation types (commercial aircraft data, smartphone pressure observations) provide more data to initialize models.

Frequently Asked Questions

Which weather model should I use for tomorrow versus next week?

Match the model to the lead time, because no single model covers both well. For weather happening today or tomorrow, the HRRR is the primary tool: 3 km resolution over CONUS, a fresh run every hour, forecasts to 18 hours on standard cycles and 48 hours at 00, 06, 12 and 18 UTC, and radar data assimilated into the initialization. A separate HRRR domain covers Alaska. In the two-to-five-day range the regional guidance only reaches partway, since the NAM stops at 84 hours (and after the scheduled October 6, 2026 transition, RRFS runs hourly out to 18 hours with longer forecasts four times a day), so compare whatever regional output covers your event against the GFS and the ECMWF, then lean on the globals alone beyond that. For next week, the GFS reaches 16 days and the ECMWF 15, and the ensemble spread matters more than any single run.

The honest caveat is that "best" is situational. Verification varies by season, region, and parameter, so a model that wins on summer convection timing may lose on winter temperatures in the same location.

If the GFS and the European model disagree, which one should I believe?

Neither one automatically. Compare current observations, official forecast discussions, and ensemble guidance for the region and weather variable you care about. A historical model reputation does not establish which forecast is right today. If two models show different arrival times, consider both scenarios rather than assuming the midpoint will be correct.

Disagreement is information in itself. Two models splitting on a storm track is a signal that the atmosphere is in a low-confidence state, which is exactly when you plan for more than one outcome rather than picking a favorite. Official forecast discussions can help explain which scenario forecasters favor and why.

What replaces the NAM after October 6, 2026?

The Rapid Refresh Forecast System (RRFS), with the RRFS Ensemble Forecast System (REFS) taking over the ensemble side. At 12 UTC on October 6, 2026, NCEP is scheduled to retire the NAM, SREF, HREF, HiresW, and NAM MOS together. RRFS is a 3-km convection-allowing model covering North America with a CONUS focus, updated hourly out to 18 hours with longer runs four times a day, replacing NAM and HiresW products. HRRR continues operating and contributes members to REFS. REFS takes over probabilistic products from HREF and SREF, which is what severe-weather forecasters actually lean on.

If your workflow depends on NAM soundings, HREF probability fields, or SREF plumes, start looking at the RRFS and REFS parallel output before the switch so you have a feel for their biases rather than learning them during a severe weather event.

Does higher resolution mean a more accurate forecast?

No. Higher resolution buys detail, not guaranteed accuracy. A 3-km grid can represent smaller storm-scale features than a 13-km grid, but the extra detail does not guarantee correct timing or placement. Extended HRRR runs reach 48 hours; the NAM 3-km nest reaches 60 hours while its 12-km parent reaches 84 hours. GFS reaches 16 days and the ECMWF 00/12 UTC control forecasts reach 15 days, and convection-allowing models can simulate storms that never happen. Fine grid spacing does not remove uncertainty in the initial conditions or model physics.

Read a high-resolution run as one plausible realization rather than a schedule. A 3-km model painting a storm over your town at 4 PM two days out is telling you the environment supports storms in that area at that time, not that the cell will be there.

How often does each model update, and how far out does each one go?

The GFS runs every 6 hours (00z, 06z, 12z, 18z) at 13 km globally out to 16 days, and the ECMWF runs every 6 hours at 9 km globally, out to 15 days from 00z and 12z and out to 6 days from 06z and 18z. The HRRR runs every hour at 3 km over the continental US, out to 18 hours on most cycles and out to 48 hours on the 00, 06, 12 and 18 UTC runs, and its coarser parent the RAP runs hourly at 13 km across North America to 21 hours on standard cycles and 51 hours at 03, 09, 15 and 21 UTC. The NAM runs every 6 hours over North America out to 84 hours, with a 3-km nest over the continental US covering the shorter lead times, until its scheduled retirement on October 6, 2026; after the confirmed transition RRFS runs hourly at 3 km out to 18 hours, with runs out to 84 hours four times a day.

These are current operational configurations, not permanent ones. NCEP and ECMWF upgrade resolution, physics, and run frequency periodically, and the October 2026 NAM retirement is a reminder that the model list itself changes.

What do ensemble forecasts add that a single model run does not?

They tell you how confident to be. A single run is one possible evolution of a chaotic atmosphere; an ensemble samples different initial conditions and may also vary boundary conditions, physics, or models to produce a range of outcomes. The GEFS carries 31 members and the ECMWF EPS 51, while on the short-range side SREF and HREF are scheduled to give way to the convection-allowing REFS on October 6, 2026. When members cluster tightly, confidence may be higher, but shared errors can still make the whole ensemble miss the outcome. When they scatter, the forecast is genuinely uncertain, and the spread is as meaningful as the mean.

Members can share errors even when they use different physics. HREF and REFS include differing model configurations, so identical physics is not a requirement for an ensemble. That is part of why forecasters compare ensembles from different centers rather than trusting one system's spread alone.

Making the Most of Model Data

No single model is always best. Understanding which model to trust, and when, separates casual forecast consumers from weather-savvy individuals.

Check multiple models and look for consensus. Model agreement can support confidence when it also fits observations and official guidance. When they diverge, acknowledge the uncertainty and prepare for multiple scenarios. Use high-resolution models for near-term details and global models for the bigger picture.

WeatherAI Forecast blends guidance from multiple models and observations. Other selectable providers use their own forecasting systems. But when you want to dive deeper, for a storm chase, a critical outdoor event, or just pure weather enthusiasm, knowing your models gives you an edge.

Now you know what the GFS, HRRR, NAM, and Euro are doing behind the scenes when you check tomorrow's forecast.

Sources and Configuration Checks

Model specifications were checked on September 6, 2026. Forecast range depends on the product and cycle, and the October transition remains subject to the NWS service notices.

• NCEP GFS products
• NCEP HRRR products, CONUS and Alaska
• NCEP RAP products and extended cycles
• NCEP NAM products
• ECMWF open data and forecast cycles
• NWS RRFS and REFS implementation notice