HiRO-ACEv2: Ai2 Put 10.9 GB of Climate Emulator Weights Online and Told Nobody

HiRO-ACEv2 is not a language model, and that is the first thing to get straight about it. It is a two-stage climate emulation framework from the Allen Institute for AI, and on 12 October 2026 at 16:43 UTC the institute created a Hugging Face repository under the name allenai/HiRO-ACEv2 containing roughly 10.9 GB of checkpoints, forcing data and initial conditions. Two days later, on 14 October at 13:15 UTC, the repository took its last commit. As of this writing it has a single like, no downloads, no discussion thread, no announcement post from Ai2, no blog entry and no press coverage. The only public explanation of what the release contains is a preprint that appeared on arXiv on 13 October, one day after the weights.

What makes this worth a page rather than a shrug is the delta from the first generation. Ai2 shipped HiRO-ACE in January 2026, and it did one thing: 3 km precipitation. HiRO-ACEv2 adds 10-metre eastward and northward winds and mean sea level pressure to the output, on top of precipitation, which moves it from a rainfall downscaler to something closer to a usable atmospheric field generator for the 3 km scale. That is a real capability change, not a version bump. Whether it holds up is a separate question, and one nobody outside Ai2 has answered yet.

What is in the repository, file by file

This is the part that is fully knowable, so it is the part worth reading carefully. Everything below was read off the repository's own file listing and metadata.

• HiROv2.ckpt — 1,376,737,727 bytes, about 1.28 GiB. This is the diffusion downscaling model itself, the component the paper calls HiRO for "High Resolution Output."

• ACE2S-X-SHiELD.ckpt — 3,201,380,831 bytes, about 2.98 GiB. The coarse-resolution emulator that feeds it: an autoregressive model of X-SHIELD, the global storm-resolving model, coarsened to 100 km.

• forcing_data/ — ten NetCDF files, forcing_2014.nc through forcing_2023.nc, roughly 609 MB each. Nine hundred and thirty megabytes of that window is conditioning input, and its ten-year span is exactly the training window both checkpoints are limited to.

• initial_conditions/ — ic_2014.nc and ic_2023.nc, 118,253,505 bytes each. Two endpoints of the period the release is set up to run.

• Three YAML configs and a shell script — ace2s_inference_config_global.yaml, hiro_downscaling_ace2s_global_output.yaml, hiro_downscaling_ace2s_pnw_output.yaml and run-hiro-ace.sh. The Pacific Northwest config sits alongside the global one, which is a fair hint about which region the authors used as their worked example.

The README is 5,345 bytes and is mostly a recipe: download the repo, pip install fme, point four fields in the ACE config at your paths, point the HiRO config at the ACE run's output directory, run the shell script. The code lives in Ai2's public climate-emulator repository and the documentation site currently advertises the package as fme v2026.5.1, with a standing warning that the codebase supports internal research and that breaking API changes happen. Weights are Apache 2.0, with Ai2's Responsible Use Guidelines attaching research and educational intent to the licence grant.

What the preprint claims, and what the model card will not say for it

The preprint is titled "Extreme events and statistics in an AI emulator of a global storm-resolving model," and its first version landed on arXiv on 13 October 2026 at 04:18 UTC under physics.ao-ph, with ten authors from Ai2's climate team. Almost every number a reader would want comes from it rather than from the model card, and every one of those numbers is the authors' own and unreproduced by anyone else.

The headline claim is cost: the paper states that inference is more than 300 times cheaper than running a conventional numerical global storm-resolving model. That is the kind of figure that decides whether a technique is a research curiosity or an operational tool, so it is worth being precise about what it is. It is a comparison of one AI emulator against one class of numerical simulation, made by the people who built the emulator, on hardware and settings they chose. No third party has repeated it.

The second claim is about tails. HiRO-ACEv2 is reported to track reference precipitation and near-surface wind distributions out to the 99.9999th percentile, and to broadly follow the reference tropical cyclone intensity distribution. Getting the far tail right is the hard part of this problem and the reason the paper exists at all — an emulator that nails the mean and misses the extremes is useless for exactly the questions people want to ask about tropical cyclones and flooding. The mechanism the authors describe is multiple denoisers at inference time, plus changes to data sampling, noise sampling and loss weighting during training.

What the model card does not do is give a benchmark number. There is no table, no score, no leaderboard entry. The card makes qualitative claims — the 99.9999th percentile line is the only quantitative figure in it — and then lists limitations with unusual candour. Both checkpoints were trained on nine years of X-SHIELD, 2014 through 2022. The outputs reproduce X-SHIELD, not observations, so they inherit the reference simulation's biases rather than the real atmosphere's. The framework is not for weather forecasting. It is not tested outside 66°S to 70°N and cannot run beyond 88°S or 88°N. Downscaling regions larger than the 16-by-16, one-degree training patch can produce discontinuities or blending artifacts at patch boundaries. And the card notes that ACE2S-X-SHIELD is less rigorously tested than the earlier ACE2-ERA5 line.

Read together, those two paragraphs say something specific: this is a research artifact with a paper behind it, released with its limitations attached, and not a product launch with a benchmark table. That is a normal and respectable way to ship science. It is also a different thing from a model you can evaluate against a public leaderboard, and the gap between those two is the honest state of this release.

The delta from v1 is the actual story here

HiRO-ACE, the January 2026 release, generated 6-hourly averaged surface precipitation rates at 3 km. That was the whole output surface. The v2 framework adds instantaneous 10-metre eastward and northward winds and sea level pressure fields at 6-hourly frequency, alongside precipitation.

That matters because of what the second stage is. HiROv2 is a diffusion model that takes 100 km inputs and produces 3 km fields; the paper describes it as a two-stage system where an autoregressive emulator generates coarsened X-SHIELD at 100 km, and a separately trained diffusion model regionally downscales that output to 3 km. Extending the downscaler from a single scalar field to a multivariate one — precipitation plus two wind components plus pressure, which have to be physically consistent with each other, not independently plausible — is the substantive engineering in this release. Wind and precipitation that disagree with each other are worse than either alone.

The v1 repository has accumulated sixteen likes since January. The v2 repository has one, two days after its last commit. That difference is not a comment on quality; it is a snapshot of how quietly this particular release went out.

What is not knowable from the outside, and probably will not be for a while

There is no hosted endpoint. There is no API, no rate card, no per-token price, and no possibility of one, because this is not an autoregressive text model. There is nothing on any model leaderboard, and nothing likely to be, since the evaluation that matters here is distributional fidelity to a reference simulation and no public leaderboard measures that. There is no independent reproduction, no third-party benchmark, and no framework integration beyond Ai2's own Python package — the configs reference the fme library and the ai2cm/ace repository and nothing else.

The practical consequence is that running HiRO-ACEv2 today means downloading 10.9 GB and executing it on your own hardware. There is no call you can make. Anyone writing that this model is "available on" some platform is describing a download link.

So the honest summary of the state of this release is: the weights exist, the recipe exists, the paper exists, and the vendor has said nothing. If Ai2 follows the pattern of the January release, an announcement may follow within days and change nothing about the artifact. If it does not, the model stays as it is — sitting in a repository with a complete README and no audience.

Where a routing layer does and does not help here

Being clear about this is more useful than stretching a product mention over it: OrcaRouter has nothing to route for HiRO-ACEv2. We do not host it, it does not appear in our catalogue, and it could not appear there in its current form, because there is no token-metered endpoint to put in front of a caller.

What is true is that the work around a model like this one is language-model work. The scripts that drive a multi-decade ensemble, the code that reads NetCDF metadata and sanity-checks field shapes before a run, the summarising of distribution statistics across a hundred rollouts, the reading of a paper into a checklist — that layer is ordinary text and code generation, and it is the layer where calling a model per request makes sense. That layer is what an API key is for, and it is why a single key in front of two hundred-plus models, with provider list price passed through and failover handled automatically when a provider degrades, is worth having even when the science model at the centre of the pipeline sits on your own GPU.

What would change this page

Three things, in descending order of likelihood. An Ai2 announcement would confirm the release is finished rather than a staging commit, and would give the framework a citation target other than the preprint. A run by somebody outside Ai2 — one group reproducing the 300-times inference claim on their own hardware, or publishing a comparison against a numerical GSRM on the same machine — would turn the paper's headline from claim into evidence. And a downstream pipeline that ingests the multivariate output, winds and pressure alongside precipitation, would be the first real test of whether the second stage's physical consistency holds up when someone other than its authors is depending on it.

None of those is a reason to ignore the release. A mature field shipping its second generation quietly is normal, and the paper is the place to look. It is a reason to be precise about what has been shown, which so far is: the files are there, the recipe is documented, and the numbers are the authors' own.

A release this quiet is also the moment to check what has gone wrong before: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.