What is HiRO-ACEv2?
HiRO-ACEv2 is an open-weight climate emulator developed by the Allen Institute for AI. It is part of the Ai2 Climate Emulator, or ACE, model collection and is designed to approximate selected outputs of a high-resolution climate model at lower computational cost. The system is aimed at scientific and educational use, not at conversational interaction or general-purpose machine learning applications.
The name describes a two-stage system. The first stage uses ACE2S-X-SHiELD, a coarse-resolution emulator that works at approximately 100 km resolution. The second stage uses HiROv2, a diffusion-based downscaling model, to transform the coarse result into climate fields at approximately 3 km resolution. In practical terms, HiRO-ACEv2 combines broad global or regional climate evolution with a learned procedure for adding finer spatial detail.
According to the supplied model information, HiRO-ACEv2 was released on October 14, 2026, and is currently described as an open-weight research model. Its model repository is available through Hugging Face under an Apache 2.0 license. No hosted inference provider or recurring API price is published for the system.
How the two-stage system works
Running a climate simulation at very fine resolution can require substantial computing resources. HiRO-ACEv2 addresses this by separating the task into two related problems. The coarse emulator represents large-scale climate behavior at roughly 100 km resolution. The downscaling stage then produces finer regional fields at about 3 km resolution.
This arrangement is useful because many climate questions require local detail that a coarse grid cannot represent well. A 3 km field can provide a more detailed view of regional precipitation and wind structures than a 100 km field. The model is therefore positioned for experiments involving high-resolution climate patterns, extreme events, and regional analysis rather than as a drop-in replacement for every type of numerical weather or climate model.
The system's documented output variables include six-hourly averaged surface precipitation and six-hour-frequency instantaneous 10 m eastward wind, 10 m northward wind, and sea-level pressure. These are numerical geophysical fields, not prose explanations or conventional AI-generated media. The model's outputs should consequently be interpreted and validated using climate-science workflows rather than language-model evaluation methods.
Primary use cases
HiRO-ACEv2 is most relevant when the objective is to study high-resolution climate behavior or generate many simulated fields for analysis. The documented target applications include:
- Extreme-weather research: examining the structure and statistics of intense events represented by the model's learned climate domain.
- Tropical-cyclone analysis: studying simulated wind, pressure, and precipitation patterns associated with cyclone-scale phenomena.
- Regional downscaling: converting coarser approximately 100 km inputs into fields with approximately 3 km spatial resolution.
- Ensemble generation: producing multiple simulated outcomes for research experiments and distributional analysis.
- Climate education and prototyping: exploring climate-emulation methods without treating the system as an operational forecast service.
Its strongest practical distinction is the combination of a global or regional climate-emulation workflow with fine-resolution output. That makes it more suitable for research questions involving spatial detail than a model that only produces coarse climate fields. However, the usefulness of the output depends on whether the target variables, forcing conditions, geographic area, and time scales match the conditions represented in training.
Inputs, outputs, and supported modalities
HiRO-ACEv2 is not a text, vision, audio, or multimodal assistant. The supplied specifications mark text input and output, image input and output, audio input and output, video input and output, music output, embeddings, structured output, speech output, and action output as unsupported or not applicable. It should be understood as a numerical climate-field generation system.
The documented outputs are climate variables on spatial grids. Specifically, the model can generate six-hourly averaged surface precipitation together with six-hour-frequency instantaneous 10 m eastward wind, 10 m northward wind, and sea-level pressure. This distinction matters: the system does not return a natural-language forecast by default, and its output variables should not be confused with a public weather report.
No conventional context-window length or maximum output-token limit is published. Those fields are not meaningful in the same way they are for a language model. Runtime behavior instead depends on the geographic area, patch size, input data, model configuration, and available compute. The supplied notes identify documented geographic and patch-size limits but do not provide a single universal token-style maximum.
Geographic and training-data limits
HiRO-ACEv2 was trained on approximately nine to ten years of X-SHiELD climate-model data. This means its learned behavior is tied to the climate-model data and forcing conditions represented in that training set. It is not a guarantee that the system will remain reliable for conditions substantially outside that domain.
The documented training coverage extends approximately from 66°S to 70°N. Runtime limits are described as extending between approximately 88°S and 88°N, with additional restrictions associated with geographic coverage and patch size. These ranges should not be interpreted as proof that every location inside them will have identical quality. They indicate the operational and training boundaries supplied for the model, while regional validity still requires scientific assessment.
A particularly important limitation is that HiRO-ACEv2 is not intended for weather forecasting. It is also not intended to emulate observational data, to generalize freely beyond its training conditions, or to serve as a general-purpose language or multimodal model. A researcher needing an operational forecast, an observation-reconstruction system, or a model calibrated for another climate regime should consider a more appropriate specialized option.
Strengths and trade-offs
The main strength of HiRO-ACEv2 is its two-stage resolution strategy. It targets approximately 3 km climate fields while using a coarse emulator as part of the pipeline. For research involving regional structure, extreme events, or ensembles, that can be more useful than working only with a coarse grid.
Its open-weight distribution is another practical advantage for researchers who need to inspect, adapt, or self-host the system. The Apache 2.0 repository license provides a permissive basis for using the released model materials, subject to the license and the project's own documentation. Self-hosting also avoids dependence on a hosted inference endpoint, although it shifts responsibility for hardware, software setup, storage, and reproducibility to the user.
The trade-off is specialization. HiRO-ACEv2 does not provide the general interface, broad task coverage, or simple per-request pricing associated with commercial language or multimodal APIs. It also has no published hosted API price, reasoning score, coding score, speed score, or cost score. Any assessment of compute cost or throughput therefore depends on the researcher's hardware and implementation rather than on a provider-published service tier.
The diffusion-based downscaling stage may be valuable for producing detailed fields, but the resulting detail should not automatically be treated as observed truth. A high-resolution output is still a model-generated representation constrained by the training data and architecture. Scientific validation remains necessary before using the outputs for substantive conclusions.
Pricing and availability
HiRO-ACEv2 has no published hosted API price for input or output. The supplied model record describes it as intended for self-hosted inference and identifies the Hugging Face repository as the available model location. Users should therefore plan around the cost of their own compute environment rather than a per-token, per-image, or subscription billing model.
There is also no documented provider-hosted streaming service, batch API, tool-use interface, JSON mode, or caching feature for this item. These are not merely undisclosed conveniences to assume; the supplied specifications mark them as unsupported or unavailable. The normal usage pattern is consequently closer to running a research model locally or on research infrastructure than calling a general-purpose cloud AI endpoint.
When to choose HiRO-ACEv2
Choose HiRO-ACEv2 when you need an open-weight climate emulator focused on high-resolution numerical fields, especially for extreme-weather research, tropical-cyclone analysis, regional downscaling, or ensemble experiments. It is a particularly relevant option when approximately 3 km output is more useful than a purely coarse climate representation and when your input conditions fall within the documented training and geographic domain.
It may be a poor choice when you need an operational weather forecast, observational-data emulation, a general-purpose AI assistant, text generation, image generation, or a hosted API with predictable per-request pricing. It is also a poor fit when the target climate regime or forcing conditions are far outside those represented by the X-SHiELD training data. In those cases, a forecasting system, an observation-focused emulator, a different regional climate model, or a commercial inference service may be more appropriate depending on the task.
For intermediate users, the key selection question is not whether HiRO-ACEv2 is broadly capable. It is whether the scientific problem matches its variables, resolution, geographic coverage, training conditions, and non-forecasting purpose. When those conditions align, its two-stage design offers a practical route to fine-resolution climate experimentation. When they do not, the apparent spatial detail can create a misleading sense of suitability.
Bottom line
HiRO-ACEv2 is a specialized Allen Institute for AI research system for generating approximately 3 km climate fields from approximately 100 km inputs. Its focus is numerical climate emulation, not language, media generation, or operational forecasting. The model's open-weight availability, two-stage architecture, and emphasis on extreme-weather and regional analysis make it useful for appropriately scoped research. Its lack of a hosted API, unspecified universal context or output limits, dependence on self-hosted compute, and restricted training domain are equally important when evaluating whether it belongs in a scientific workflow.