Today, we're introducing PEAKS 2.0, our latest state-of-the-art weather prediction engine that now drives all OpenSnow forecasts for the next 15 days.
PEAKS 2.0 marks a significant upgrade over PEAKS 1.0 and other weather prediction systems.
PEAKS 2.0 predicts temperature, wind, and precipitation for the continental United States and for various international regions.
Why does this matter for you?
Not all forecasts are created equal.
In an age when anyone can vibe code a weather app using freely (or cheaply) available weather forecast data, part of our differentiation is to create forecasts that are uniquely better than other free products.
Part of being "better" is to start with more accurate forecasts, and our testing (see below) shows that PEAKS 2.0 is more accurate than other forecast models and the National Weather Service.
- PEAKS 2.0 more accurately predicts future weather because it learned from actual (not modeled) past weather at thousands of stations.
- PEAKS 2.0 is more detailed because it can 'see' valleys, ridges, and slopes, with a resolution of 1.5 miles and adjustments down to a hyperlocal 50 feet.
- PEAKS 2.0 is more helpful because it shows a range of outcomes (rarely is a single model correct).
- PEAKS 2.0 is more useful because it extends its high-resolution predictions out to 15 days (most high-resolution models stop at 2-4 days).
Accuracy
Compared to other models and forecast systems, PEAKS 2.0 consistently scores first place across all variables.
In our evaluations across the continental United States, PEAKS 2.0 forecasts outperform NOAA's National Blend of Models (NBM), the National Weather Service (NWS) forecasts, as well as every individual model evaluated (European ECMWF, American GFS, Canadian GDPS, German DWD-ICON, and more).
Of 10 million temperature forecasts:
Of 8 million wind forecasts:
Of 2 million precipitation forecasts:
Here is how we evaluate the forecasts.
Average Miss (mean absolute error):
- For a given forecast, how many degrees or miles-per-hour or inches is the forecast different from the actual observation?
Significant Error Rate:
- How often is the model off enough to matter?
- For temperature, how often is the model off by more than 3 degrees Fahrenheit; for wind, how often is the model off by more than 3 miles per hour; and for precipitation, how often does the model miss measurable precipitation, predict measurable precipitation when there is none, or predict precipitation with an error greater than 0.5 inches?
Most Often Best:
- How often is the model the most accurate of the evaluated models?
- How often, over the entire forecast period and across all of the observation sites, is the model the best or tied with the best. For precipitation, those probabilities are all pretty high because most models accurately predict no precipitation on obviously dry days.
On each of these evaluation metrics, PEAKS 2.0 scores the best.
How does it work?
PEAKS 2.0 is a machine learning model trained to take a baseline weather model forecast, at a low resolution, and emulate it at a much higher resolution.
The point forecast system refines the forecast from a low resolution of ~15 mile boxes, to the emulated resolution of ~1.5 miles, to a hyper-local ~50 foot resolution point.
Instead of the forecast changing every, ~15 miles, the forecast changes every ~50 feet based on the local terrain.
Also, built-in bias correction allows the model to learn from past mistakes using actual weather station data.
One common mistake (bias) in models is the inversion bias - models underestimate cold pools in valleys on calm, clear nights, leading to temperature forecasts that are too warm in the valleys.
For example, this is a common occurrence in Steamboat due to the local terrain.
PEAKS 2.0 attempts to fix this mistake (bias) at Steamboat and anywhere that its learned temperature inversions exist.
Where is PEAKS 2.0 available?
PEAKS 2.0 is now available in the shaded regions below.
In the OpenSnow app, select any location, and scroll down to the "Forecast Range" and "Precipitation by Model" views to compare PEAKS 2.0 against other models.
Why is PEAKS 2.0 different from other models?
Conventional weather models take a current state of the atmosphere and simulate how the weather will change over time.
- Models' granularity is defined by its horizontal resolution.
- Most global models run at a resolution of 5-15 miles, meaning that the model is calculating what will occur at every 5-15 mile grid box.
- Some higher-resolution models run at ~1.5 miles, but are only available in smaller regions and at shorter ranges (2-4 days into the future). Examples are the US-focused HRRR and the Europe-focused ICON-D2.
PEAKS 2.0 is inherently different from any of those models.
- Instead of taking an estimated state of the atmosphere and simulating it forward using physics equations or trained learnings from models, PEAKS 2.0 uses actual weather stations to learn what is actually occurring at the ground.
- PEAKS 2.0 is a sophisticated model that consists of a combination of downscaling (bringing global data from its low resolution to a much higher resolution) and bias correction (learning mistakes/biases from historical data).
What about PEAKS 1.0?
PEAKS 1.0 was our first effort to build a more accurate weather prediction system.
- PEAKS 1.0 used historical estimated/modeled data to correct mistakes/biases.
- PEAKS 2.0 uses historical actual station data to correct mistakes/biases.
For training, actual weather station data is better than estimated/modeled data, and this is why PEAKS 2.0 is an upgrade from PEAKS 1.0.
State-of-the-Art Technology
PEAKS 2.0 combines elements of deep learning, weather prediction model emulation, bias correction, and model weighting to provide accurate forecasts for exact locations.
We are proud to say that as of October 15, 2026, PEAKS 2.0 drives the temperature, precipitation, snowfall, wind speed, wind direction, and wind gust elements of OpenSnow forecasts (other models still contribute to our multi-model forecast, but the multi-model forecast is heavily weighted toward PEAKS 2.0).
We will continue to train, refine,
and release new versions of
PEAKS.
As the oft-repeated saying goes, this version of PEAKS should be the worst it ever is, or said another way, each version of PEAKS should get better and better.
Questions?
Send an email to hello@opensnow.com — a real human will get back to you within
24 hours.