Ai2's HiRO-ACEv2 generates global 3 km climate data at about 1/300th the cost of X-SHiELD, it claims

Ai2 has released HiRO-ACEv2, an update to its High Resolution Output Ai2 Climate Emulator that, according to the company, generates global 3 km climate data at about 1/300th the cost of X-SHiELD, a physics-based global atmospheric model that explicitly simulates storms and other small-scale weather. The update was published October 14, 2026.

How the system works

HiRO-ACE, first introduced earlier in 2026, has two components:

• ACE: an AI model that runs at a coarse 100 km resolution.
• HiRO: a downscaling model that sharpens ACE's coarse output to 3 km detail.

The main change in v2 is the new downscaling model, HiROv2. It makes two upgrades over v1:

1. It predicts several high-resolution fields at once, rather than only rainfall. The fields covered are rain, wind and pressure.
2. It has improved skill on extreme events, which Ai2 says was the main goal of the release. Changes to the training procedure improved reproduction of the most intense rain, wind and pressure values.

Ai2 also trained an updated version of ACE for the system.

Reported results

All figures below come from Ai2's own evaluation against X-SHiELD, the simulation HiRO was trained on.

• Extreme-event accuracy: given coarsened X-SHiELD data, HiROv2 matches the target's high-resolution statistics to within 3%, including events that occur once in a million data points.
• Tropical cyclones: the model accurately estimates total cyclone energy, including for storms absent from training. It generates hurricanes with realistic eyes and rain bands. Ai2 acknowledges that winds and pressure around the eye are still somewhat more lopsided than in X-SHiELD.
• Terrain effects: HiROv2 realistically generates strong winds flowing down mountain slopes.
• Long runs: over a 10-year simulation, the updated ACE's hurricanes, average climate patterns, and distributions of rain, wind and pressure line up well with X-SHiELD.
• Full pipeline: the combined system produced a full year of global 3 km data whose statistics closely match X-SHiELD's. HiRO worked from a simulation ACE ran on its own rather than from real X-SHiELD data. Across a 10-year run, hurricane energy is comparable to X-SHiELD's.

Cost

Ai2 says global high-resolution output costs roughly 300 times less than running X-SHiELD. For a single region, it claims savings can be more than 100 times greater still, depending on the area's size. It is also developing a streamlined HiROv2 that it says could be more than 10 times faster again. No compute hardware or dollar cost figures were disclosed.

Limitations

Because HiRO learned from X-SHiELD, it works only with the ACE version tuned to match X-SHiELD. It cannot yet produce high-resolution data for other climates, such as a warmer future or historical records like ERA5. Ai2's planned next steps:

• Train on new X-SHiELD datasets, including one simulating a world 4°C warmer.
• Add outputs such as near-surface temperature and snow and ice, which would enable uses like snowpack prediction.

Access

The models are available to download, and Ai2 provides tutorial notebooks, a tech report and a web demo. The demo shows pre-generated precipitation, wind and sea-level pressure from HiROv1 and HiROv2 beside X-SHiELD, and users can zoom into a region of interest. Parameter count, license terms and training cutoff were not stated in the source material.

What this means

The 3% statistical agreement on one-in-a-million events is the most important claim here, because tail behavior is what matters for climate risk work. Those results are in-distribution, though. HiRO is validated against the same X-SHiELD simulation that produced its training data, and the system cannot yet represent other climates. Until the planned 4°C-warmer training set arrives, HiRO-ACEv2 is best seen as a cheap way to generate large ensembles of X-SHiELD-like extremes, not a general projection tool for future climate. If the cost claims hold up under independent replication, researchers could run far more ensemble members for extreme-event statistics than 3 km physics models allow. The remaining hurricane-eye asymmetry shows that fine-scale fidelity still trails the physics-based reference.