What this category does
AI weather models learn the behavior of the atmosphere from decades of past data, then produce forecasts far faster than traditional physics-based models that solve equations on a supercomputer. Companies in this category either sell forecasts from their own model, sell access to the model so you can run it yourself, or build products on top of models that others have published.
The category is young, and the offers vary a great deal. Some are research systems that produce global fields at a fixed resolution. Others are operational services with service commitments, regional tuning and support. Reading the product description carefully matters more here than in older categories.
How buyers evaluate
Buyers start by asking what the model was trained on, how it is initialized, and whether it has run in real time long enough to judge. A model that scores well on a historical test set has shown something, but not what it will do next week when the data feeding it arrives late or incomplete. Look for a record of live operation and ask how the provider handles a missed input.
Evaluation usually means a side-by-side test against the forecasts you use today. Pick a period with the kind of weather that drives your decisions, such as a storm, a cold snap or a heat wave, and compare against observations. AI models can be strong on average and weak on extreme events, so an average alone will mislead you. Trials of one to three months are typical, and a full season is better if your business is seasonal.
RFP criteria tend to include resolution, lead time, update frequency, supported parameters, licensing of outputs, and the compute you need if you run the model yourself.
Be careful with headline claims in this category. A comparison published by a model's own developer usually uses a test period and a set of variables chosen by that developer. That does not make it wrong, but it does mean you should repeat the comparison on the variables and locations that your business depends on, using observations that neither party controls.
Questions to ask a vendor
- Is the model run by you in real time, or do I receive a static set of results?
- Which input data does it depend on, and what happens when that input is delayed?
- Which parameters and lead times does it cover, and which are experimental?
- Has an independent party evaluated it against observations, and may I see the method?
- If I get model access, what hardware do I need and who supports me?
- What are the terms for using the outputs in my own products?
- How do you version the model, and how will I know when it changes?
How to read the capability facts
The forecast types and parameters on a profile show what a vendor offers. The product types field is the one to read first in this category, because it separates a forecast you consume from model access you operate. Delivery tells you how results reach you, whether by an API, bulk files or an embedded SDK.
Profile facts describe an offer. They do not describe how well the model performs. For that, look for independent evaluation, and treat a Verified credential as a sign that a provider has had its forecasts measured against observations by a third party.
Ask also about the future of the offer. Models in this field are updated often, and a new version can change behavior. A vendor that tells you in advance when a version changes, and lets you keep running the old one for a time, makes it easier to hold your own results steady.
Where to go next
Use these guide questions to separate research demos from operational services before you commit time to a trial.
Browse the AI Weather Models providers, filter them by region, parameter and pricing, or compare several side by side. The full list of categories is the place to start if your need spans more than one.