It’s always exciting when a new numerical weather model becomes publicly available, and the latest arrival is no exception. The Rapid Refresh Forecast System (RRFS), nicknamed “Roofus,” officially launched on November 3, 2026, bringing a new generation of high-resolution forecasting.
In this post, we’ll explore RRFS vs. HRRR, comparing the new Rapid Refresh Forecast System with the established High-Resolution Rapid Refresh (HRRRv4). We’ll examine several meteorologically interesting variables and highlight where the two models agree and differ.
RRFS vs. HRRR: What Are We Comparing?
This RRFS vs. HRRR comparison is more of a lemons-to-limes exercise than an apples-to-apples one. However, it’s still an interesting look at what each model produces. We aren’t claiming one model is better than the other. Independent verification through observation campaigns will help determine how each performs. Instead, we’ll focus on what the models are producing and what their differences can tell us.
Each figure shows one run from each model at the same valid time, alongside a difference panel (RRFS minus HRRR). Each difference panel includes two statistics:
- Mean (μ): The average difference across the frame. A positive value means RRFS is higher overall.
- Standard deviation (σ): Measures how much the difference varies from place to place. A mean near zero with a large standard deviation means the models agree on average but may disagree locally.
Both models were processed with Boxer, our cloud-native weather data processing system. We currently serve HRRR, with RRFS coming soon.
RRFS vs. HRRR: Comparing Weather Variables
Few meteorological variables relate more to human comfort than the wet-bulb globe temperature (WBGT). The WBGT is a standardized way to quantify how we feel (assuming we are small, hollow, rubbery black spheres).
We have a whole blog post on this, so we’ll defer the interested reader there. Across the country, the two models nearly cancel each other out (μ = 0.01 °F), yet σ = 1.11 °F shows local differences of a degree or more.
Next up is plain vanilla 2 m temperature. The average difference is again small (μ = −0.15 °F), but the spread is wider (σ = 1.71 °F) than for WBGT.
Now let’s zoom in on the El Paso region to take a closer look, starting with the surface (10 m) wind.
In this frame, RRFS winds are lighter on average (μ = −0.51 m/s), but σ = 0.77 m/s is larger than that offset, so the sign flips from place to place.
Wind results from air pressure trying to smooth itself out. Let’s look at mean sea level pressure differences over the El Paso region.
The land generally sits above sea level, so each model must estimate the pressure beneath it using their best estimate of the topography at 3 km resolution. Much of the difference follows the terrain. Here μ = 1.33 hPa is larger than σ = 1.19 hPa, so the difference is mostly one sign: an offset more than scatter.
Now for precipitation, shown as composite reflectivity near Salt Lake City, Utah. The two models place the echoes in the same broad area but differ in the details of individual cells. Treat μ = −7.47 dBZ with caution, as there is no precipitation over some regions, so the two models report different background values whose offset dominates the average.
Taking a Closer Look at Model Differences
This RRFS vs. HRRR comparison also highlights how differences in model calculations can affect derived variables like dewpoint. Zooming in on dewpoint, RRFS is the more humid model, at least in this instantaneous comparison. The mean difference (μ) is 1.87 °F and the standard deviation (σ) is 3.12 °F. Because σ is well above μ, that average hides pockets where HRRR is more humid. The largest gaps across this single forecast hour in the region exceed 8 °F.
What’s Different Under the Hood?
RRFS has a longer forecast horizon (84 hours versus 48) and broader spatial coverage (North America versus the contiguous U.S. and Alaska). Both models provide data at 3 km resolution.
However, they are built on different dynamical cores, the engines that advance the governing equations of the atmosphere through time. RRFS uses FV3, while HRRR uses WRF-ARW. Derived fields, such as dewpoint, are also calculated differently in each model.
What’s Next for RRFS and HRRR?
One notable addition in RRFS is the 1.5 km fire weather domain, which runs out to 36 hours.
RRFSv2 promises another shake-up in conventional numerical weather modeling. The dynamical core itself is expected to change from FV3 to MPAS, which uses a hexagonal Voronoi mesh. As for HRRR, NOAA has no major upgrades planned, and a future version of RRFS is expected to eventually retire it, possibly as soon as 2028.
Until RRFSv2 arrives, we have two high-resolution 3 km models to explore. Lemons and limes aren’t the same fruit, but they’re both worth having in the kitchen.
Bringing Weather Models to Life
All the examples in this post came from gridded weather data processed with Boxer. But processing the data is only part of the equation. Making it accessible and useful through visualization is just as important.
Our Terrier visualization toolkit can turn those same grids into interactive map layers. Here’s an example using HRRR composite reflectivity:
This RRFS vs. HRRR comparison shows that even two high-resolution weather models covering similar areas can produce meaningful forecast differences. As new models like RRFS become available, the ability to process, access, and visualize high-resolution weather data becomes increasingly important. We’re looking forward to exploring what RRFS brings to the table and making that data available through the Wet Dog Weather platform.
