3 min readData and AI

MITRE Weather 1K: the 1 km AI training dataset and AceCAST's part

In April 2026 MITRE and The Weather Company announced Weather 1K, a 1 km AI training dataset. The announcement names AceCAST among the modeling tools used to develop it.

Illustration: North America at night from orbit, with city lights and cloud systems traced by glowing lines of airflow.

First published May 7, 2026 · Updated November 16, 2026

Teams that train AI weather models hit the same wall early: a model can only learn the detail that is in its training data. Many public archives are coarser than a single thunderstorm and are saved once an hour. On April 23, 2026, MITRE and The Weather Company announced a collaboration built around a dataset made to close that gap, called Weather 1K. The announcement names TempoQuest's AceCAST among the modeling tools used to develop it. This article sets out what was announced, where AceCAST fits, and what to do if your team needs high-resolution training data of its own.

What MITRE and The Weather Company announced

The two organizations announced the collaboration on April 23, 2026. Weather 1K is a high-resolution AI training dataset. The Weather Company licensed it for research and development on its GRAF forecasting system.

The announcement describes the dataset in four numbers:

  • Resolution: 1 km spatial resolution.
  • Coverage: the continental United States, from sea level to 62,000 ft.
  • Cadence: 10-minute increments.
  • Size: roughly 7 PB.

The numbers explain each other. A 1 km grid over the continental United States has millions of columns. Each column is a stack of levels reaching up to 62,000 ft, well above the heights where airliners cruise. Every 10 minutes the whole stack is saved again. That is how a weather dataset reaches petabytes.

Where AceCAST fits

The announcement names TempoQuest's AceCAST among the modeling tools used to develop Weather 1K. Weather 1K is MITRE's dataset, and the announcement does not detail how each tool was used.

AceCAST runs WRF, the Weather Research and Forecasting model maintained by NCAR, on NVIDIA GPUs. It keeps the WRF workflow, with WRF-compatible inputs, namelists and outputs, and it runs on Linux systems with NVIDIA GPUs, on premises or in the cloud. The AceCAST documentation has the detail.

A physics-based model such as WRF solves the equations of the atmosphere on a grid. Its output is complete: every variable, at every grid point, at every saved time. That completeness is what makes model output usable as training data in places and at heights where observations are sparse.

TempoQuest published its own press release about the announcement on May 6, 2026.

Why AI weather models need data like this

An AI weather model learns the patterns in its training set and nothing else. Three properties of that set decide what the trained model can do.

  • Grid spacing. If the training fields are too coarse to hold a thunderstorm, the model has never seen one in detail and cannot draw one.
  • Cadence. Storms grow and decay in tens of minutes. Hourly snapshots miss most of that life cycle. Fields saved every 10 minutes capture it.
  • Consistency. In physics-model output, wind, temperature, moisture and pressure obey the same physical relationships everywhere. The AI model learns those relationships from the data.

One caution applies to every dataset of this kind. Model output is not observation. A dataset inherits the biases of the model and the settings that produced it, so a careful team scores it against stations, radar or other measurements before training on it.

If your team needs a dataset of its own

TempoQuest does not resell Weather 1K. What TempoQuest offers is separate: Custom Weather Data, scoped for each customer.

For a custom project, TempoQuest scopes, generates, packages and delivers project-specific high-resolution model outputs and derived datasets for AI training, risk analytics, planning and operations. The scope covers:

  • region and variables
  • resolution (typical requests are around 1 km or 3 km)
  • cadence (15-minute, 30-minute, hourly or daily)
  • history and forecast horizon, including historical datasets
  • format and delivery (NetCDF, GRIB, Zarr, CSV, JSON or an API)

Pricing follows the work: domain size, compute, history, cadence, validation and integration.

Before you ask for a quote, settle four questions. They fix most of the scope and most of the cost.

  1. What decision or model will the data feed, and over what area?
  2. How fine and how frequent does it need to be? Finer grids and shorter intervals both multiply compute and storage.
  3. How many years of history do you need to capture enough of the rare events you care about?
  4. Which observations will you score the data against?

A team that arrives with those answers gets a scope, and a price, much sooner. If you already run WRF and want to generate the data yourself, AceCAST runs it on NVIDIA GPUs, and a trial is the way to test it on your own domain.

What to do next

Scope your dataset, or see custom weather data.

Your training dataset

Bring your data requirements.

Share the region, archive, fields and rights your model needs. We review every request.