Ai2 releases HiRO-ACEv2 open climate weights for 3 km extreme-weather simulation

The Allen Institute for AI has released HiRO-ACEv2, an updated open model system for generating high-resolution climate simulations. Announced on October 14, the system pairs ACE, a coarse atmospheric emulator operating around 100-kilometer resolution, with HiROv2, a diffusion-based downscaler that generates approximately 3-kilometer fields. The model repository and associated tutorials provide a practical path for researchers to inspect and run the published checkpoints rather than relying solely on a research announcement.

The upgrade extends HiRO beyond precipitation alone. Ai2 says the new downscaler generates precipitation, near-surface eastward and northward winds, and sea-level pressure, allowing research into storms and other severe weather patterns. The team emphasizes statistical representation of rare extremes and tropical cyclones. Its comparison target is X-SHiELD, a computationally intensive high-resolution atmospheric simulation. This is climate emulation of a particular simulation system, not a direct replacement for observations or a guaranteed weather prediction service.

Ai2 reports that the new downscaler can match certain high-resolution extreme-event statistics within three percent when supplied with coarsened X-SHiELD inputs. For a full global simulation, the institute estimates a cost around one three-hundredth of X-SHiELD. Those are research-team measurements under defined conditions, not independently replicated Open Model Weights benchmarks. A less expensive emulator also does not establish accuracy in every geographic region, climate scenario or operational forecasting workflow.

Important restrictions are documented. Ai2 says HiROv2 remains tied to an ACE configuration trained to match X-SHiELD and does not yet support arbitrary historical or warmer-future climates. The model card also describes geographic coverage and patch-boundary constraints. Hugging Face labels the release Apache-2.0 and the publisher provides model files and inference instructions; exact artifact and code dependencies still deserve review. For the OMW registry, HiRO-ACEv2 is a scientific model intake candidate requiring pinned checkpoints, license files, model lineage and field-specific provenance, not a reason to infer broader general-purpose language-model capabilities.