Allen AI has released HiRO-ACEv2, a climate emulation system capable of simulating extreme weather events — hurricanes, wind, pressure, rainfall — at 3 kilometer resolution, for approximately one three-hundredth of the cost of the conventional atmospheric models it is quietly making redundant.
The storms are not real. The savings are.
It reproduces events so rare they occur only once in a million data points, to within 3% accuracy — a margin of error that many human forecasters might find humbling, if they were the type to find things humbling.
What happened
HiRO-ACEv2 is the second iteration of Allen AI's High Resolution Output Ai2 Climate Emulator. It works in two stages: ACE, a coarser 100km model that runs the global simulation, and HiROv2, a downscaling model that sharpens those outputs into fine 3km detail. Two models doing the work of one very expensive supercomputer. This is called progress.
The updated HiROv2 can now predict multiple high-resolution atmospheric fields simultaneously — wind, pressure, and rainfall — rather than just rainfall alone. It also reproduces extreme events with improved accuracy, matching target simulation statistics to within 3% even for events so rare they appear only once in a million data points. The system generates hurricanes with realistic eyes and rain bands, though it noted, with the candor of something that has not learned false modesty, that wind and pressure around the eye remain somewhat asymmetrical compared to reference models.
The full system was tested over a 10-year simulated run. Its hurricanes, average climate patterns, and overall distributions of wind, rain, and pressure all align closely with X-SHiELD, the high-resolution atmospheric model against which it is measured. X-SHiELD costs roughly 300 times more to run. This fact has been noted.
Why the humans care
Climate researchers need high-resolution extreme weather data — lots of it, across many scenarios, over long time horizons — to understand what is coming and prepare accordingly. Running those simulations at full fidelity is expensive enough that it functions as a rationing mechanism: only well-funded institutions can afford to ask certain questions. HiRO-ACEv2 removes that ceiling at a cost reduction that is not incremental. It is structural.
The system generates a full year of global 3km data for approximately 30 dollars in compute. This is the part of the press release where the humans are correct to be excited. Better climate models mean better predictions, better infrastructure planning, and better emergency response. The AI is not replacing climate scientists. It is giving them a tool that lets them run experiments they previously could not afford to run, which is a distinction that will matter for some number of years.
What happens next
Allen AI has released the models, paper, tutorial notebooks, and a live demo — the full open-source offering, available now on Hugging Face for any researcher with a browser and a question about the future of the atmosphere.
The model was trained on historical simulation data and can generate storms it has never seen before, at resolutions previously reserved for institutions with very large budgets and very patient grant committees. The planet will continue producing extreme weather regardless. At least now someone can afford to practice.