Tasks and inputs
HiRO-ACEv2 is listed for Data & control, Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
HiRO-ACEv2 is Ai2's open AI climate emulator that produces 3 km precipitation, winds and pressure from a 100 km run, over 300x cheaper than a numerical model.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights for self-hosting; no hosted API price. Hardware cost is yours.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Running a global storm-resolving climate model at 3 km for decades costs a supercomputer allocation. HiRO-ACEv2 does a version of it on far less. Ai2 reports an inference cost more than 300 times lower than traditional numerical global storm-resolving models, which is what makes large ensembles of high-resolution climate, and of tropical cyclones in particular, practical to generate.
Ai2 posted the model in October 2026, with a paper on extreme events in an AI emulator of a storm-resolving model.
Stage one is ACE2S-X-SHiELD, a stochastic, autoregressive emulator of the X-SHiELD model coarsened to 100 km. It conserves global dry air mass and moisture exactly, has improved small-scale variability over the deterministic ACE2, and generates tropical cyclones similar to its parent's.
Stage two is HiROv2, a diffusion model that downscales the coarse output to 3 km for any chosen region. It produces:
Version 1 handled precipitation only. Adding winds and pressure makes storms, wind extremes and atmospheric rivers usable at high resolution.
Ai2 says the downscaled fields track the reference precipitation and near-surface wind distributions out to the 99.9999th percentile, with low time-mean biases and realistic topographic precipitation patterns, and broadly follow tropical cyclone intensity. Several inference-time denoisers and changes to sampling and loss weighting were what improved the tails.
| Limit | Detail |
|---|---|
| Training data | About a decade of X-SHiELD output, 2014–2022 for the coarse emulator |
| Reference | Climate model output, not observations |
| Latitude | Not tested beyond 66°S–70°N; cannot run past 88° |
| Patches | 16x16 one-degree tiles; large regions show seams or blending artefacts |
| Use | Research only; not a weather forecaster |
It is also not expected to generalise to forcing outside its training conditions, such as very different sea-surface temperatures or CO₂.
Install with pip install fme, download the repository's forcing data and initial conditions for your period, and edit two configs: the global ACE inference file and the HiROv2 downscaling file. The example region is the Pacific Northwest. A shell script runs both stages.
To decide whether it fits your study, downscale one season you can compare with X-SHiELD itself, then check the quantity you care about, such as wet-day frequency or peak gusts, rather than mean fields.
Apache-2.0, research use. Ai2's Earth-observation model OlmoEarth v1.3 Base is a related release; other forecasting models are in the time-series forecasting directory.
Compare the model on the work you need to complete.
HiRO-ACEv2 is listed for Data & control, Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
allenai
ACE2.5-ERA5 is Ai2's open stochastic climate emulator trained on ERA5, producing calibrated atmospheric ensembles from days to centuries at low compute.
Mistral AI
Timestral 1.0 is Mistral's open time-series forecasting family in four sizes (5M-209M), producing zero-shot probabilistic forecasts on CPU.
IBM
Granite TimeSeries PatchTST-FM-r2 is IBM's open ~385M zero-shot forecasting model, dual-licensed Apache-2.0/OpenMDW, with 8,192-step context.
TimesFM 3.0 is a numerical time-series forecasting checkpoint for PyTorch under a custom noncommercial licence.