About Indra
Indra forecasts US airports with Indra-3M, a weather model with 3.0 million parameters in a 12 MB file. For each of the next 72 hours, it predicts temperature, dew point, wind and pressure, with a range around each forecast. This site shows that next to the National Weather Service forecast for the same place.
It runs on your device. Your browser downloads the model once, fetches the public weather data itself, and computes the forecast right there. Nothing about the forecast runs on a server of ours.
What it reads
- The airport's last 90 days of hourly measurements: temperature, dew point, wind, pressure, gusts, cloud cover, rain and visibility.
- Three forecast sources, for those past days and the next 72 hours: the National Blend of Models, and the GFS and NAM model output statistics.
- The local solar time, from the airport's longitude, and the day of year.
Because it sees each source's recent forecasts next to what actually happened at that airport, it can adjust for things like "the Blend has been running 2° cold here this week". That's why one small model can work across the country.
How it was made
It started as Timestral-5M, a time-series model released by Mistral AI. We fine-tuned it on 24 US airports plus Indianapolis, using August 2018 to October 2024. At every point it saw only data that had actually been published by then.
We then pruned it from 5.2 million parameters to 3.0 million over 12 rounds. Each round measured, on training airports only, how much of each neuron's output the others couldn't reproduce and how much the forecast error depended on what was left. It removed whatever cost least, rewrote the weights so the remaining neurons covered for it, and retrained. A larger 28M model acted as a teacher during training, but isn't part of what runs here.
We used November 2024 to October 2025, at Indianapolis and six airports it didn't train on, to choose between versions.
For each hour, the model outputs 21 points of a probability distribution, not a single number. The line on the forecast page is the middle of that distribution, and "Likely ranges" shows the 50% and 80% ranges around it.
Track record
Final test: November 2025 to October 2026, about 1,470 forecasts per airport, scored once on the finished model. It never trained on these six airports. Average temperature error, 1–72 hours ahead, in °F:
| Airport | Indra-3M | National Blend | Lower by |
|---|---|---|---|
| Columbus | 2.04 | 2.27 | 10.2% |
| Chicago O'Hare | 2.08 | 2.31 | 10.2% |
| Denver | 2.79 | 3.29 | 15.4% |
| Miami | 1.42 | 1.59 | 10.3% |
| Seattle | 1.61 | 1.79 | 9.7% |
| Phoenix | 1.70 | 2.03 | 15.9% |
We had used those six airports to check earlier models, so we also wrote down 12 airports no model here had ever seen, then scored them once, with 240 forecasts each over the same year: Buffalo, Burlington, Raleigh-Durham, Savannah, Mobile, Austin, El Paso, Wichita, Fargo, Billings, Grand Junction and Sacramento. Its temperature error was 11.7% lower than the Blend's overall, and between 3.7% (Buffalo) and 15.4% (Billings) by airport.
We only have a history of official NWS forecasts at Indianapolis. On identical hours over that same year, average error by hours ahead:
The model trained on Indianapolis through October 2024, but not on this year.
Where it works
The forecast page covers 2,407 airports in the lower 48 states, the ones the National Blend covers. Search for a city and you get a nearby airport, usually the closest large one, and how far away it is. The forecast is for that airport's sensor, so a city miles away can differ, especially on calm, clear nights.
We've scored or trained the model at 49 of those airports. The rest haven't been scored, and small airports without GFS or NAM forecasts may do worse. The page says so when that applies. Alaska, Hawaii and other US territories aren't included, because it never trained on anything like them.
Rain
The hourly rain forecast is separate from Indra-3M, only covers Indianapolis, and is computed ahead of time, not on your device. Elsewhere, the rain line is the NWS forecast. The rain model starts from the National Blend's rain chances and learns how to correct them. For each hour, it gives the chance of at least 0.01 inches of rain or melted snow at the airport.
Here is how it scored over the same test year, using the Brier score, which measures how good a percentage is. Lower is better.
| Hours ahead | Ours | Blend, 6 h chance scaled to one hour | HRRR, rain or no rain |
|---|---|---|---|
| 1–12 h | 0.029 | 0.037 | 0.045 |
| 13–24 h | 0.033 | 0.038 | 0.057 |
| 25–48 h | 0.036 | 0.040 | 0.065 |
| 49–72 h | 0.041 | 0.042 | no data |
Its percentages can also be taken at face value. Over the same year, 1 to 12 hours ahead:
| When it said | It rained |
|---|---|
| 20–30% | 23% of the time |
| 30–50% | 38% of the time |
| 50–70% | 56% of the time |
| 70–100% | 80% of the time |
Neither comparison above is an hourly forecast built to compete with ours, so this shows the model is useful, not that it's the best available. We have no history of the NWS's hourly rain chances to score against. We started saving them on November 4, 2026, so that comparison starts from there.
The Blend's rain chances come in 6-hour windows, so ours can jump where one window ends: at 1 and 7, morning and evening, during standard time. A sudden drop at one of those hours says less about exactly when the rain will stop.
How often it updates
Each forecast is computed when you open the page, from the newest data published, and again every 30 minutes or so while it stays open. If a data source is late, the forecast uses older runs or goes without.
What it doesn't do (yet)
- Anything past 72 hours.
- Anywhere but airports. Downtown is often a few degrees warmer on calm, clear nights.
- Safety decisions. It's a research model, so don't use it for aviation or emergencies.