All stories
HiRO-ACEv2 Launches Source-backed

Ai2 says HiRO-ACEv2 makes 3 km climate simulations 300 times cheaper

The downloadable system adds wind and pressure to high-resolution rainfall output, with cost and accuracy claims measured against a physical climate simulation.

Sources (2)

Why now Ai2 announced HiRO-ACEv2 on October 14 at 16:12 UTC; the model repository was created October 12, so today's announcement is distinct from its initial upload.

Wind and pressure join the rainfall model

Ai2 has announced HiRO-ACEv2, a climate simulation system it says can generate global data at 3 km resolution for roughly one three-hundredth of the computational cost of X-SHiELD, the physical atmospheric model used as its reference. Its new HiROv2 downscaler generates rain, wind and sea-level pressure together; the earlier version focused on rainfall. Public checkpoints and example configurations are available under Apache 2.0.

The system starts with ACE’s coarser, approximately 100 km atmospheric output and uses a diffusion model to add fine-scale structure. Ai2 reports improved reproduction of extreme rain, wind and pressure values and comparable tropical-cyclone energy across a ten-year simulation. Those are comparisons with X-SHiELD output, not an independently established improvement in real-world weather forecasts. The announcement also acknowledges uneven wind and pressure patterns around simulated hurricane eyes.

A research simulator with a defined climate boundary

The model card explicitly says the included ACE checkpoint is unsuitable for weather forecasting. HiROv2 was trained on X-SHiELD simulations, and the current framework cannot yet generate high-resolution output for a warmer future climate or the historical climate represented by ERA5. Broader climate training and additional outputs remain planned work.

Researchers can inspect pre-generated output in Ai2’s visualizer or follow the downloadable model’s fme inference instructions. The visualizer is a way to compare simulated fields, rather than a live forecast service.