Skip to main content

Pluvium

Pluvium is a rain model. It says when each parish might get rain, down to the exact town or city, and the time when rain is going to fall.

I made it for Jamaica first. Then I did not want it mainly focusing on Jamaica, other countries should be able to use it too, so now you give it a country code and it sets up that country by itself. Any country between 60 south and 60 north works.

You don't need a GPU for the forecast and you don't need an account. You copy one command, you run it, it should just work. If it has to download something it downloads it once and keeps it, it does not redownload the stuff every time because I find that completely useless.

pip install pluvium
pluvium now --country JM

Is it a standalone model

Is Pluvium a standalone model, or did I use another model to power it? Both. The 14 days forecast is not standalone, that one uses the forecasts ECMWF and NOAA put out free every day, six of them, some are the AI weather models. Then it looks at one year of rain from the satellite for that country to see which of the six was right more times, and those it trusts more. That is the blend. No GPU for that part, any normal computer runs it.

The one that is standalone is Pluvium-500, a specific model that forecasts rain hour by hour. It uses the NatureV1 architecture and some extended features. It only works after it is trained on a GPU.

The six forecasts From Kind
ENS, 51 runs ECMWF physics
HRES ECMWF physics
AIFS ECMWF AI
GEFS, 31 runs NOAA physics
GFS NOAA physics
GraphCast, called AIGFS since May 2026 NOAA AI

Can it tell the time when it is going to rain

Can it tell the time when it is going to rain? For the first three days yes, but in 6 hour parts and not the exact hour. In Jamaica the parts are 7 pm to 1 am, 1 am to 7 am, 7 am to 1 pm and 1 pm to 7 pm. The exact hour is what Pluvium-500 is for and that one needs the GPU training first.

If it says 34% for the afternoon then most times no rain is going to fall, and thats correct, not a mistake. Also places that are close get almost the same forecast, it works in squares of about 28 km.

How good is it

Test Result
Rain or no rain, days 1 to 10, six regions, one full year best of all the systems in 6 of 6 regions
Same, days 11 to 14 best in 5 of 6
Against each system on its own, days 1 to 7 better in 72 of 72 comparisons, 95% confidence
50 mm days, version 0.3, seven places ahead of every other system in 15 of 28 cases
Winter, days 1 to 3 ahead in only 2 of 6 regions
Hurricane Melissa in Jamaica, two days ahead lost: said 66% for 50 mm, the raw systems said 83 to 100%
Part of the day, Jamaica, wet days 7.8% better than the usual pattern
Part of the day, picking the wettest one no better: 66.7% against 66.5%
Hurricane tracks, East Pacific 2025, 18 storms lower error than ECMWF HRES and NOAA GFS at all 8 lead times
Hurricane tracks against the official NHC forecast behind at every lead time, 125 against 100 nautical miles at five days

Is it finished

Pluvium-500 Status
Training cell notebooks/pluvium_everything.py, needs a Hugging Face token and a big GPU
GPU it was made for RTX 6000 Blackwell, 96 GB
GPU memory used 95% of the card each step
Finished no, first run stopped at 138 minutes
Hour by hour forecast not published yet

What it can not do

Can not do Why
The exact hour only 6 hour parts until Pluvium-500 is trained
Two places closer than 28 km same forecast square
Rain gauge accuracy checked against satellite rain only
Learn from last week satellite rain is published 4.5 months late
Much past day 10 only a little better than the usual weather
50 mm days hard for every forecast tested

More commands

Command What it gives
pluvium now --country JM the biggest towns
pluvium now --country JM --place "Port Antonio" one place by name
pluvium now --country JM --when the 4 parts of each day, first 3 days
pluvium now --country JM --parishes every parish
pluvium now --country JM --parish "St Ann" the towns of one parish
pluvium now --country DO,HT --storms two countries, and the hurricanes alive now

More

Page What is in it
docs/GUIDE.md the long manual
docs/TRAINING.md the training commands
CHANGELOG.md every version

Apache 2.0, CorX Labs.

Metadata

Release files for pluvium 0.5.5

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for pluvium 0.5.5
File Size Uploaded Provenance Actions
pluvium-0.5.5.tar.gz 1.5 MB Without provenance Details for pluvium-0.5.5.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for pluvium 0.5.5
File Interpreter ABI Platform Actions
pluvium-0.5.5-py3-none-any.whl Python 3 none any Details for pluvium-0.5.5-py3-none-any.whl

Total release size: 2.1 MB

Release files / pluvium-0.5.5.tar.gz

Download URL pluvium-0.5.5.tar.gz
Size 1.5 MB
Tags Source
SHA-256 checksum
How to use checksums
6a2287022f09bccaf7133c5ab54fd6105975e3318a41fef49a6f22c24336aa19
BLAKE2b-256 checksum
How to use checksums
048b5f4db08e69ff70e38a5b12260a33cb8cb1affc064ad183986ec2b571c4f6
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.18

Release files / pluvium-0.5.5-py3-none-any.whl

Download URL pluvium-0.5.5-py3-none-any.whl
Size 519.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
a475e89cdb6bc04219d2dd4a89a512c1e50139b8af7dc4cc87ef787fe10f2f6f
BLAKE2b-256 checksum
How to use checksums
8b898e64e7d740994f9dea2e1f5bd936d6db6dc6bf84c1f0387a3b3dc9c08575
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.18

Security for pluvium 0.5.5

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page