HiRO-ACEv2 can't yet generate 3 km storm data for a warmer world, even with Ai2's reported 300-fold cut in inference cost. Released on October 14, the system learns from the physics-based X-SHiELD atmospheric model. Developers using it for climate-risk tools need to work within that simulated climate.
The model card rules out weather forecasting and generalization beyond the training conditions for sea-surface temperature and carbon dioxide. Its detailed storm maps are samples from a modeled climate.
Why cheaper storm samples matter
The original HiRO-ACE announcement in January described the computational problem: global atmospheric simulations with grid spacing around 3 km can represent storms and interactions with terrain that coarse, roughly 100 km climate models miss. Running those fine grids for a decade can take months. That makes repeated experiments expensive before researchers have even begun comparing rare events.
HiRO-ACE uses two probabilistic components. ACE produces coarse atmospheric trajectories, then HiRO generates finer regional outcomes from them. Different samples let researchers examine a range of possible storm structures. The January version generated high-resolution precipitation, recovering details such as rainfall differences over mountainous terrain. The approach already separated the inexpensive global simulation from the more costly local detail. Ai2's explanation puts that division at the center of the design.
Version two adds winds 10 meters above the surface and sea-level pressure alongside rainfall, with improved representation of extremes. Those fields expand the range of storm characteristics researchers can examine. They are the substantive additions in the new release.
A six-hour rainfall average isn't a wind snapshot
The release's model card describes HiRO as a diffusion model that converts 100 km inputs into 3 km output. Its rainfall values are averages over six-hour intervals. Wind components and pressure are instantaneous values at six-hour spacing. Anyone combining the fields in an analysis needs to preserve that distinction.
The supplied regional configuration saves rainfall, both horizontal wind components and pressure. It includes a decade of output for one region and a separate single-event example with 16 ensemble members. The example's comment warns that the named event may not exist in your data. A timestamp in this stochastic simulation doesn't establish that a corresponding real-world storm happened then.
You can study repeated versions of one simulated event or generate a longer regional record. The downscaling documentation supports both individual snapshots and time ranges, with separate controls for the number of ensemble members and GPU batch size. That lets a developer limit the experiment before committing to a much larger output.
Accuracy is measured against a simulation
The released training configuration uses simulated data from 2014 through 2022, with 2023 reserved for validation. Those dates describe the reference simulation's calendar. They don't turn its storms into an observational record. Holding out a year tests reproduction on unseen simulation output, within the climate represented by that dataset.
Training also gives patches centered between 30 degrees south and 30 degrees north eight times the sampling weight of other regions. The low-noise denoiser assigns rainfall a loss multiplier of 0.2. These choices are visible in the configuration, along with wind and pressure output channels. They show deliberate tuning of the sampling and training objective. Anyone adapting the method to another dataset would need to assess those choices again.
With coarsened X-SHiELD inputs, Ai2 reports differences within 3% at the target distributions' 99.9999th percentile, reaching extremes exceeded by roughly one in a million data points. The target is the high-resolution simulation. That result in the announcement measures reproduction of a modeled distribution, rather than accuracy against observed weather.
The technical report adds a caution for applications sensitive to individual storms: the strongest cyclone winds tend to be underpredicted, and independently downscaled snapshots have less continuity through time. Matching an aggregate distribution therefore leaves room for errors in a storm's peak or evolution.
HiROv2's model card limits tested coverage to 66 degrees south through 70 degrees north. Larger regions can show discontinuities or blending artifacts where generated patches meet. Fine grid spacing doesn't remove those restrictions, especially for analysis crossing patch boundaries.
Energy savings and elapsed time differ
The paper reports global inference energy use at roughly 0.3% of X-SHiELD's. Its wall-clock comparison is about 33 times faster, using eight NVIDIA B200 GPUs against X-SHiELD on 27,648 CPU cores. These are different measurements on different hardware configurations. The 300-fold figure describes inference cost. It doesn't mean every job finishes 300 times sooner. The cost section gives the comparison.
The regional numbers are more practical for a first experiment. Downscaling one simulated year over a 16-by-16-degree patch takes about 12 minutes on a B200, 15 on an H100 or 30 on an L40S. Downscaling a global year takes roughly 2.75 days on one B200. These are the researchers' reported timings, not measurements from our own run. Their benchmark makes the cost of selecting a region visible.
Selecting a region also controls the computational budget. The documentation warns that leaving spatial bounds unspecified generates a very large dataset. It also rules out generation below the model's training patch size. You can reduce the area you request, within the supported patch geometry, instead of treating global output as the default starting point.
Checkpoints and analysis code are available
Ai2's ACE repository contains the Python package fme for training, running and evaluating its atmospheric emulators, plus data-processing scripts and model configurations. The distinction between the software family and this checkpoint matters: ACE covers several models, while HiRO-ACEv2 ships a particular pairing with its own constraints.
The published model package includes both checkpoints, forcing data and initial conditions, with configurations for global and regional downscaling. Its card gives the release an Apache 2.0 license. You can inspect the experiment's inputs alongside its generated storm images.
For the output side, the inference documentation specifies separate Zarr files with time, ensemble, latitude and longitude dimensions. It exposes variable selection as well as spatial bounds. A first run can therefore answer a bounded question about selected fields without generating every available output. Multiple output requests run sequentially, while generation can be parallelized across GPUs.
Ai2's workshop notebooks include a HiRO-ACE walkthrough and were adapted for free T4 GPUs on Google Colaboratory. They offer a way to learn the two-stage workflow. The workshop predates this release, so its hardware requirements and examples shouldn't be confused with the new paper's global inference benchmark.
The manuscript repository contains configuration files and analysis and plotting scripts, and links to an archive of intermediate plotting data. Researchers can check those materials when a result depends on preprocessing or how the comparison was calculated.
Warmer-climate checkpoints are still planned
The current model's documented weaknesses include uncertainty over whether runs much longer than its nine training years yield independent, representative samples. Cheaper generation doesn't by itself establish that a longer synthetic record contains more useful information. Researchers need to check that assumption before treating extra simulated decades as extra evidence.
Ai2 plans training with additional X-SHiELD datasets, including a climate 4°C warmer, and intends to add temperature and frozen-precipitation outputs. Those are future tasks in the release announcement. The next result to watch is whether new checkpoints reproduce extremes in those climates. Users could then assess warming-related risks against the limits demonstrated in those evaluations.