Ai2's HiRO-ACEv2 downscales storms to 3 km — but only for one climate model's world
The Allen Institute for AI (Ai2) has released HiRO-ACEv2, an open-weights climate emulator that, according to the institute itself, can generate global climate data at 3 km resolution for roughly one three-hundredth of the cost of X-SHiELD, a high-resolution global atmosphere model that explicitly simulates storms and other small-scale weather. The numbers are Ai2's own and, so far, independently unverified — and the system is trained on data from only a single climate model.
What was launched, and when
The Hugging Face repository allenai/HiRO-ACEv2 was created on 12 October 2026 at 16:43 UTC and received its last commit on 14 October at 13:15 UTC, according to a review of the repository's metadata in OrcaRouter's blog post on the release. Ai2's own announcement on Hugging Face, published under the Ai2Comms account with Kyle Wiggers listed, is dated 14 October, while an arXiv preprint followed on 13 October.
There is a timing ambiguity worth noting: OrcaRouter described the release as unannounced, with zero downloads, when its blog post was written, while Ai2's announcement is also dated 14 October. The exact sequence between the two is unclear, but what stands firm is that both the weights and a detailed explanation of the system are now publicly available. The weights are licensed under Apache 2.0.
How the system works
HiRO-ACEv2 is the update to Ai2's High Resolution Output Ai2 Climate Emulator (HiRO-ACE) and consists of two stages:
- The ACE2S-X-SHiELD emulator: an AI model trained on X-SHiELD outputs that can generate global atmospheric simulations quickly.
- HiROv2, a new diffusion-based downscaling model: the component that makes it possible to go from coarser global simulations to 3 km resolution — fine enough to capture storms and other small-scale weather.
The biggest change in this version is precisely HiROv2, which according to Ai2's announcement can now "predict multiple high-resolution fields simultaneously instead of only precipitation", allowing it to capture a much broader range of extreme weather — rain, wind and pressure — while having improved accuracy on extreme events. The previous version handled precipitation only.
The numbers — and what they cover
All the key figures come from Ai2 itself and have not yet been independently verified:
- Cost: According to Ai2, the full system can generate global 3 km climate data for roughly one three-hundredth of the cost of X-SHiELD.
- Accuracy: When HiROv2 is given coarse X-SHiELD data, Ai2 says it matches the target simulation's high-resolution statistics within 3%, "even for events so rare that they occur only once per million data points".
- Hurricane energy: Over a 10-year run, the energy of HiRO-ACEv2's hurricanes is, according to Ai2, comparable to X-SHiELD's over the same period.
- Regional savings: For region-only downscaling, Ai2 claims the savings can exceed 100 times beyond the global 300x factor.
These claims come from Ai2's announcement and the accompanying preprint; no external evaluation exists yet. OrcaRouter itself points out that no one outside Ai2 has responded on whether the capability holds up.
What's in the repository?
According to OrcaRouter, which read the file listing directly from the repository, the release contains roughly 10.9 GB spread across:
- HiROv2.ckpt — 1,376,737,727 bytes (approx. 1.28 GiB), the diffusion downscaling model itself.
- ACE2S-X-SHiELD.ckpt — approx. 2.98 GiB, the emulator checkpoint.
- Ten NetCDF forcing files covering the years 2014–2023.
- Initial condition files.
- Configurations, including one for downscaling the Pacific Northwest region.
All of it under the Apache 2.0 license, which permits both commercial use and redistribution.
The limitations Ai2 itself points out
The most significant limitation is that the entire system is tied to the ACE version trained on X-SHiELD. That means, as Ai2 writes in the announcement, that it "cannot yet produce high-resolution data for other climates, such as a warmer future scenario or the historical record in datasets like ERA5".
As the next step, Ai2 plans to train on new X-SHiELD datasets, "including one that simulates a world 4 °C warmer, so that the system can handle different climates".
For anyone hoping to use HiRO-ACEv2 for future scenarios or to downscale historical observational data, this is a significant constraint: in practice, the system can currently only generate 3 km data for the climate world that X-SHiELD represents.
Open questions
- Independent verification: Neither the 300x cost saving nor the 3% accuracy has been confirmed by anyone outside Ai2. Since the entire system only works on X-SHiELD data, a natural test would be to compare HiRO-ACEv2's outputs against original X-SHiELD runs.
- Adoption: OrcaRouter observed the repository with zero downloads and one like at the time of writing; current usage is unknown.
- The preprint: The technical details of the 13 October arXiv preprint have not been reviewed here; the claims above are known via Ai2's announcement and OrcaRouter's review.
- Extension to other climates: The planned 4 °C-warmer training data will determine whether the system can be used for climate scenarios that decision-makers actually care about.
The release is thus a concrete, openly available tool with highly ambitious self-reported numbers — and with a validation process that currently rests entirely with Ai2 itself.