HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

Hugging Face has introduced HiRO-ACEv2, the newest update to its High Resolution Output Ai2 Climate Emulator (HiRO-ACE) system. The original HiRO-ACE, introduced earlier in 2026, pairs two models: ACE, which simulates global climate at a coarse 100 km resolution, and HiRO, which sharpens those coarse outputs into fine 3 km detail. The release is accompanied by a paper, tutorial notebooks, downloadable models and a web demo.
The central change in HiRO-ACEv2 is a new downscaling model, HiROv2. Where the earlier version predicted only rainfall at high resolution, HiROv2 predicts several high-resolution fields at once, which lets it capture a much wider range of extreme weather. Training changes also improved its skill at reproducing the most intense rain, wind and pressure values.
The headline result is accuracy on rare events. When given coarsened X-SHiELD data — X-SHiELD being a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather — HiROv2 matches the target simulation's high-resolution statistics to within 3 percent, even for events so rare they occur only once in a million data points. It also estimates the total energy of tropical cyclones accurately, including storms it never saw during training. Generated hurricanes show realistic eyes and rain bands, although winds and pressure around the eye are somewhat more lopsided than in X-SHiELD. The model additionally produces strong winds flowing down mountain slopes in a realistic way.
Hugging Face also trained an updated ACE for the system. Over a 10-year simulation, its hurricanes, average climate patterns and overall distributions of rain, wind and pressure all line up well with X-SHiELD, which the team presents as evidence that ACE's coarse simulations are a solid foundation for long-term high-resolution output.
Combining HiROv2 with ACE, the team generated a full year of global 3 km data. Its statistics closely match X-SHiELD's even though HiRO was working from a simulation ACE ran on its own rather than from actual X-SHiELD data. Across a 10-year run, the energy of HiRO-ACEv2's hurricanes is comparable to X-SHiELD's over the same period, and snapshots show a variety of other extreme events resembling those in X-SHiELD.
The practical appeal is cost and flexibility. Producing global high-resolution data with HiRO-ACEv2 costs roughly 300 times less than running X-SHiELD. When only one region is needed, the savings can be more than a hundred times greater depending on the area's size. The team says it is working on a streamlined version of HiROv2 that could make it more than ten times faster still. The system is positioned as a way for researchers to generate many versions of high-resolution rain, wind and pressure data for extreme events over 10 years or more.
To make the outputs easy to inspect without running any inference, the team built a web demo that shows pre-generated precipitation, wind and sea-level pressure from HiROv1 and HiROv2 side by side with X-SHiELD, the simulation they were trained on. Users can zoom into their own region of interest to judge performance there. For deeper analysis, the models can be downloaded and tutorial notebooks walk through how to run them. The demo complements quantitative comparisons in the technical report.
There are clear limits. Because HiRO learned from X-SHiELD, it currently works only with the version of ACE tuned to match X-SHiELD, so it cannot yet produce high-resolution data for other climates such as a warmer future or historical records like ERA5. The team plans to train on new X-SHiELD datasets, including one simulating a world 4 degrees Celsius warmer, and to add outputs such as near-surface temperature, snow and ice, which would open uses like snowpack prediction.
This is a research system rather than an operational forecasting service. In practice, that means developers and product teams should read the cost figures as relative compute comparisons for generating simulated data, not as guarantees about forecast skill on real-world observations. A likely trade-off is that the dramatic savings come partly from the coarse ACE stage doing most of the work, so the fidelity of the final 3 km output depends on how well that coarse simulation matches the target climate — which is exactly why the current version is locked to the X-SHiELD-tuned ACE. The planned 4 degrees Celsius dataset is one route toward broader applicability, though results there are not yet reported.
The most immediate audience is climate and weather researchers who need large volumes of high-resolution extreme-event data without the expense of explicit storm-resolving simulation, plus teams exploring downscaling architectures more generally. The multi-field, multi-variable framing of HiROv2 also provides a template for emulators that must preserve statistics at the tail of a distribution, not just average behavior.
Why it matters: HiRO-ACEv2 lowers the cost of producing global 3 km climate data by roughly 300 times versus X-SHiELD, and regional runs can be more than a hundred times cheaper still, which puts large ensembles of extreme-weather scenarios within reach of smaller research groups. For developers, the combination of downloadable models, notebooks and a side-by-side visual demo makes it relatively easy to evaluate the output on a region of interest before committing compute. The main constraint to plan around is that the system is currently tied to the X-SHiELD-matched ACE version, so applying it to future or historical climates is future work rather than a supported path today — and that limitation is my reading of the stated roadmap, not a claim about performance.
Based on reporting from the original publisher. Visit the source for full context and later updates.